Tech Guides - Data

[box type="note" align="" class="" width=""]This article is an excerpt from a book by Kuntal Ganguly titled Learning Generative Adversarial Networks. The book gives a complete coverage of Generative adversarial networks. [/box] The article highlights some of the common challenges that a developer might face while using GAN models. Common challenges faced while working with GAN models Training a GAN is basically about two networks, generator G(z) and discriminator D(z) trying to race against each other and trying to reach an optimum, more specifically a nash equilibrium. The definition of nash equilibrium as per Wikipedia: (in economics and game theory) a stable state of a system involving the interaction of different participants, in which no participant can gain by a unilateral change of strategy if the strategies of the others remain unchanged. 1. Setting up failure and bad initialization If you think about it, this is exactly what a GAN is trying to do; the generator and discriminator reach a state where they cannot improve further given the other is kept unchanged. Now the setup of gradient descent is to take a step in a direction that reduces the loss measure defined on the problem—but we are by no means enforcing the networks to reach Nash equilibrium in GAN, which have non-convex objective with continuous high dimensional parameters. The networks try to take successive steps to minimize a non-convex objective and end up in an oscillating process rather than decreasing the underlying true objective. In most cases, when your discriminator attains a loss very close to zero, then right away you can figure out something is wrong with your model. But the biggest pain-point is figuring out what is wrong. Another practical thing done during the training of GAN is to purposefully make one of the networks stall or learn slower, so that the other network can catch up. And in most scenarios, it's the generator that lags behind so we usually let the discriminator wait. This might be fine to some extent, but remember that for the generator to get better, it requires a good discriminator and vice versa. Ideally the system would want both the networks to learn at a rate where both get better over time. The ideal minimum loss for the discriminator is close to 0.5— this is where the generated images are indistinguishable from the real images from the perspective of the discriminator. 2. Mode collapse One of the main failure modes with training a generative adversarial network is called mode collapse or sometimes the helvetica scenario. The basic idea is that the generator can accidentally start to produce several copies of exactly the same image, so the reason is related to the game theory setup we can think of the way that we train generative adversarial networks as first maximizing with respect to the discriminator and then minimizing with respect to the generator. If we fully maximize with respect to the discriminator before we start to minimize with respect to the generator everything works out just fine. But if we go the other way around and we minimize with respect to the generator and then maximize with respect to the discriminator, everything will actually break and the reason is that if we hold the discriminator constant it will describe a single region in space as being the point that is most likely to be real rather than fake and then the generator will choose to map all noise input values to that same most likely to be real point. 3. Problem with counting GANs can sometimes be far-sighted and fail to differentiate the number of particular objects that should occur at a location. As we can see, it gives more numbers of eyes in the head than originally present: 4. Problems with perspective GANs sometime are not capable of differentiating between front and back view and hence fail to adapt well with 3D objects while generating 2D representation from it as follows: 5. Problems with global structures GANs do not understand a holistic structure similar to problems with perspective. For example, in the bottom left image, it generates an image of a quadruple cow, that is, a cow standing on its hind legs and simultaneously on all four legs. That is definitely unrealistic and not possible in real life! It is very important when its comes to train GAN models towards the execution and there would be some common challenges that can come ahead. The major challenge that arises is the failure of the setup and also the one that is mainly faced in training GAN model is mode collapse or sometimes the helvetica scenario. It highlights some of the common problems like with counting, perspective or be global structure. The above listings are some of the major issues faced while training a GAN model. To read more on solutions with real world examples, you will need to check out this book Learning Generative Adversarial Networks.  

When MyBucks, a Luxembourg based Fintech firm, started scaling up their business in other countries. They faced a daunting challenge of reducing the timeline for processing credit requests from over a week’s time to just under few minutes. Any financial institution dealing with lending could very well relate to the nature of challenges associated with giving credit - checking credit history, tracking past fraudulent activities, and so on. This automatically makes the lending process tedious and time consuming. To add to this, MyBucks also aimed to make their entire lending process extremely simple and attractive to customers. MyBucks’ promise to its customers: No more visiting branches and seeking approvals. Simply login from your mobile phone and apply for a loan - we will handle the rest in a matter of minutes. Machine Learning has triggered a whole new segment in the Fintech industry- Automated Lending Platforms. MyBucks is one such player. Some other players in this field are OnDeck, Kabbage, and Lend up. What might appear transformational with Machine Learning in MyBucks’ case is just one of the many examples of how Machine Learning is empowering a large number of finance based companies to deliver disruptive products and services. So what makes Machine Learning so attractive to Fintech and how has Machine Learning fuel this entire industry’s phenomenal growth? Read on. Quicker and efficient credit approvals Long before Machine Learning was established in large industries unlike today, it was quite commonly used to solve fraud detection problems. This primarily involved building a self-learning model that used a training dataset to begin with and further expanding its learning based on incoming data. This way the system could distinguish a fraudulent activity from a non-fraudulent one. Modern day Machine Learning systems are no different. They use the very same predictive models that rely on segmentation algorithms and methods. Fintech companies are investing in big data analytics and machine learning algorithms to make credit approvals quicker and efficient. These systems are designed in such a way that they pull data from several sources online, develop a good understanding of transactional behaviours, purchasing patterns, and social media behavior and accordingly decide creditworthiness. Robust fraud prevention and error detection methods Machine Learning is empowering banking institutions and finance service providers to embrace artificial intelligence and combat what they fear the most-- fraudulent activities. Faster and accurate processing of transactions has always been the fundamental requirement in the finance industry. An increasing number of startups are now developing Machine Learning and Artificial Intelligence systems to combat the challenges around fraudulent transactions or even instances of incorrectly reported transactions. Billguard is one such company that uses big data analytics and makes sense of millions of consumers who report billing complaints. The AI system then builds its intelligence by using this crowd-sourced data and reports incorrect charges back to consumers thereby helping get their money back. Reinventing banking solutions with the powerful combination of APIs and Machine Learning Innovation is key to survival in the finance industry. The 2017 PwC global fintech report suggests that the incumbent finance players are worried about the advances in the Fintech industry that poses direct competition to banks. But the way ahead for banks definitely goes through Fintech that is evolving everyday. In addition to Machine Learning, ‘API’ is the other strong pillar driving innovation in Fintech. Developments in Machine Learning and AI are reinventing the traditional lending industry and APIs are acting as the bridge between classic banking problems and the future possibilities. Established banks are now taking the API (Application Programming Interface) route to tie up with innovative Fintech players in their endeavor to deliver modern solutions to customers. Fintech players are also able to reap the benefits of working with the old guard, banks, in a world where APIs have suddenly become the new common language. So what is this new equation all about? API solutions are helping bridge the gap between the old and the new - by helping collaborate in newer ways to solve traditional banking problems. This impact can be seen far and wide within this industry and Fintech as an industry isn’t just limited to lending tech and everyday banking alone. There are several verticals within the industry that now find increased impact of Machine Learning -payments, wealth management, capital markets, insurance, blockchain and now even chatbots for customer service to name a few. So where do you think this partnership is headed? Please leave your comments below and let us know.

Convolutional Neural networks (CNNs), are a group from the neural network family that has manifested in areas such as Image recognition, classification, etc. They are one of the popular neural network models present in nearly all of the image recognition tasks that provide state-of-the-art-results. However, these CNNs have drawbacks, which are to be discussed later in the article. In order to address the issue with CNNs, Geoffrey Hinton, popularly known as the Godfather of Deep Learning, recently proposed a research paper along with two other researchers, Sara Sabour and Nicholas Frosst. In this paper, they introduced CapsNet or Capsule Network--a neural network, based on multi-layer capsule system. Let’s explore the issue with CNNs and how CapsNet came as an advancement to it. What is the issue with CNNs? Convolutional Neural Network or CNNs are known to seamlessly handle image classification tasks. They are experts in learning at a granular level; where the lower layers detect edges and shape of an object, and the higher layers detect the image as a whole. However, CNNs perform poorly when an image possesses a slightly different orientation (rotation or a tilt), as it compares every image with the ones it learns during training. For instance, if an image of a face is to be detected, it checks for facial features such as nose, two eyes, mouth, eyebrows, etc; irrespective of the placement. This means CNNs may identify an incorrect face in cases where the placement of an eye and the nose is not as conventionally expected, for example in case of the profile view. So, the orientation and the spatial relationships between the objects within an image is not considered by a CNN. To make CNNs understand orientation and spatial relationships, they were trained profusely with images taken from all possible angles. Unfortunately, it resulted in excess amount of time required to train the model. Also, the performance of the CNNs did not improve largely. Pooling methods were also introduced at each layer within the CNN model for two reasons; first  to reduce the time invested in training, and second to bring out positional invariance within CNNs. It resulted in triggering false positives in an image, i.e., it detected the object within an image but did not check its orientation. Also it incorrectly declared it as a right image. Thus, positional invariance made the CNNs susceptible to minute changes in viewpoint. Instead of invariance, what CNNs require is equivariance-- a feature that makes CNNs adapt to change in rotation or proportion within an image. This equivariance feature is now possible via Capsule Network! The Solution: Capsule Network CapsNet or Capsule network is an encapsulation of nested neural network layers. Traditional neural network contains multiple layers whereas a capsule network contains multiple layers within a single capsule. CNNs go deeper in terms of height, whereas the capsule network deepens in terms of nesting or internal structure. Such a model is highly robust to geometric distortions and transformations, which are a result of non-ideal camera angles. Thus, it is able to exceptionally handle orientations, rotations and so on. CapsNet Architecture Source: https://arxiv.org/pdf/1710.09829.pdf Key Features: Layer based Squashing In a typical Convolutional Neural Network, the squashing function is added to each layer of the CNN model. A squashing function compresses the input to one of the ends of a small interval, introducing nonlinearity to the neural network and enables the network to be effective. Whereas, in a Capsule network, the squashing function is applied to the vector output of each capsule. Given below is a squashing function proposed by Hinton in his research paper. Squashing function Source: https://arxiv.org/pdf/1710.09829.pd Instead of applying non-linearity to each neuron, the squashing function applies squashing to a group of neurons i.e the capsule. To be more precise, it applies nonlinearity to the vector output of each capsule. The squashing function also tries to squash the vector output to zero if it is a small vector. If the vector is too long, the function tries to limit the output vector to 1. Dynamic Routing Dynamic routing algorithm in CapsNet replaces the scalar-output feature detectors of the CNN with the vector-output capsules. Also, the max pooling feature in CNNs, which led to positional invariance, is replaced with ‘routing by agreement’. The algorithm ensures that when they forward propagate the data, it goes to the next most relevant capsule in the layer above. Although dynamic routing adds an extra computational cost to the capsule network, it has been proved to be advantageous to the network by making it more scalable and adaptable. Training the Capsule Network The capsule network is trained using the MNIST. MNIST is a dataset which includes more than 60,000 handwritten digit images. It is used to test machine learning algorithms. The capsule model is trained for 50 epochs with a batch size of 128 parts, where each epoch is responsible for a complete run through the training dataset. A TensorFlow implementation of the CapsNet based on Hinton’s research paper is available in GitHub repository. Similarly, CapsNet can also be implemented using other deep learning frameworks such as Keras, PyTorch, MXNet, etc. CapsNet is a recent breakthrough in the field of Deep learning and have a promise to benefit organizations with accurate image recognition tasks. Also, implementations with CapsNet is slowly catching up and is expected to reach at par like CNNs. They have been trained on a very simplistic dataset i.e the MNIST. They will still require to prove themselves on various other datasets. However, as time advances and we see CapsNet being trained within different domains, it will be exciting to discern how it moulds itself as a faster and more efficient training technique for deep learning models.

We know your time is valuable. That's why what matters is important. We've written about the trends and issues that are going to matter in data science, but here you can find 5 data science tools that you need to pay attention to in 2018. Read our 5 things that matter in data science in 2018 here. 1. TensorFlow Google's TensorFlow has been one of the biggest hits of 2017 when it comes to libraries. It’s arguably done a lot to make machine learning more accessible than ever before. That means more people actually building machine learning and deep learning algorithms, and the technology moving beyond the domain of data professionals and into other fields. So, if TensorFlow has passed you by we recommend you spend some time exploring it. It might just give your skill set the boost you’re looking for. Explore TensorFlow content here. 2.Jupyter Jupyter isn’t a new tool, sure. But it’s so crucial to the way data science is done that it’s importance can’t be understated. And as pressure is placed on data scientists and analysts to communicate and share data in ways that empower stakeholders in a diverse range of roles and departments. It’s also worth mentioning its relationship with Python - we’ve seen Python go from strength to strength throughout 2017, and showing no signs of letting up; the close relationship between the two will only serve to make Jupyter more popular across the data science world. Discover Jupyter eBooks and videos here. 3. Keras In a year when deep learning has captured the imagination, it makes sense to include both libraries helping to power it. It’s a close call between Keras and TensorFlow which deep learning framework is ‘better’ - ultimately, like everything, it’s about what you’re trying to do. This post explores the difference between Keras and TensorFlow very well - the conclusion is ultimately that while TensorFlow offers more ‘control’, Keras is the library you want if you simply need to get up and running. Both libraries have had a huge impact in 2017, and we’re only going to be seeing more of them in 2018. Learn Keras. Read Deep Learning with Keras. 4. Auto SkLearn Automated machine learning is going to become incredibly important in 2018. As pressure mounts on engineers and analysts to do more with less, tools like Auto SKLearn will be vital in reducing some of the ‘manual labour’ of algorithm selection and tuning. 5. Dask This one might be a little unexpected. We know just how popular Apache Spark is when it comes to distributed and parallel computing, but Dask represents an interesting competitor that’s worth watching throughout 2018. It’s high-level API integrates exceptionally well with Python libraries like NumPy and pandas; it’s also much more lightweight than Spark, so it could be a good option if you want to avoid building out a weighty big data tech stack. Explore Dask in the latest edition of Python High Performance.

Generative Adversarial Networks (GANs) are a prominent branch of Machine learning research today. As deep neural networks require a lot of data to train on, they perform poorly if data provided is not sufficient. GANs can overcome this problem by generating new and real data, without using the tricks like data augmentation. As the application of GANs in the Machine learning industry is still at the infancy level, it is considered a highly desirable niche skill. Having an added hands-on experience raises the bar higher in the job market. It can fetch you a higher pay over your colleagues and can also be the feature that sets your resume stand apart. Source: Gartner's Hype Cycle 2017  GANs along with CNNs and RNNs are a part of the in demand deep neural network experience in the industry. Here are five reasons why you should learn GANs today and how Kuntal Ganguly’s book, Learning Generative Adversarial Networks help you do just that. Kuntal is a big data analytics engineer at Amazon Web Services. He has around 7 years of experience building large-scale, data-driven systems using big data frameworks and machine learning. He has designed, developed, and deployed several large-scale distributed applications, without any assistance. Kuntal is a seasoned author with a rich set of books ranging across the data science spectrum from machine learning, deep learning, to Generative Adversarial Networks, published under his belt.[/author] The book shows how to implement GANs in your machine learning models in a quick and easy format with plenty of real-world examples and hands-on tutorials. 1. Unsupervised Learning now a cakewalk with GANs A major challenge of unsupervised learning is the massive amount of unlabelled data one needed to work through as part of data preparation. In traditional neural networks, this labeling of data is both costly and time-consuming. A creative aspect of Deep learning is now possible using Generative Adversarial Networks. Here, the neural networks are capable of generating realistic images from the real-world datasets (such as MNIST and CIFAR). GANs provide an easy way to train the DL algorithms. This is done by slashing down the amount of data required to train the neural network models, that too, with no labeling of data required. This book uses a semi-supervised approach to solve the problem of unsupervised learning for classifying images. However, this could be easily leveraged into developer’s own problem domain. 2. GANs help you change a horse into a zebra using Image style transfer https://www.youtube.com/watch?v=9reHvktowLY Turning an apple into an orange is Magic!! GANs can do this magic, without casting a  spell. Transferring Image-to-Image style, where the styling of one image is applied to the other. What GANs can do is, they can perform image-to-image translations across various domains (such as changing apple to orange or horse to zebra) using Cycle Consistent Generative Network (Cycle GANs). Detailed examples of how to turn the image of an apple to an orange using TensorFlow, and how of turn an image of a horse into a zebra using a GAN model, are given in this book.  3. GANs inputs your text and outputs an image Generative Adversarial networks can also be utilized for text-to-image synthesis. An example of this is in generating a photo-realistic image based on a caption. To do this, a dataset of images with their associated captions are given as training data. The dataset is first encoded using a hybrid neural network called the character-level convolutional Recurrent Neural network, which creates a joint representation of both in multimodal space for both the generator and the discriminator. In this book, Kuntal showcases the technique of stacking multiple generative networks to generate realistic images from textual information using StackGANs.Further, the book goes on to explain the coupling of two generative networks, to automatically discover relationships among various domains (a relationship between shoes and handbags or actor and actress) using DiscoGANs. 4. GANs + Transfer Learning = No more model generation from scratch Source: Learning Generative Adversarial Networks Data is the basis to train any Machine learning model, scarcity of which can lead to a poorly-trained model, which can have high chances of failure. Some real-life scenarios may not have sufficient data, hardware, or resources to train bigger networks in order to achieve the desired accuracy. So, is training from scratch a must-do for training the models? A well-known technique used in deep learning that adapts an existing trained model for a similar task to the task at hand is known as Transfer Learning. This book will showcase Transfer learning using some hands-on examples. Further, a combination of both Transfer learning and GANs, to generate high-resolution realistic images with facial datasets is explained. Thus, you will also understand how to create creating artistic hallucination on images beyond GAN. 5. GANs help you take Machine Learning models to Production Most Machine learning tutorials, video courses, and books, explain the training and the evaluation of the models. But how do we take this trained model to production and put it to use and make it available to customers? In this book, the author has taken an example, i.e. developing a facial correction system using an LFW dataset, to automatically correct corrupted images using your trained GAN model. This book also contains several techniques of deploying machine learning or deep learning models in production both on data centers and clouds with micro-service based containerized environments.You will also learn the way of running deep models in a serverless environment and with managed cloud services. This article just scratches the surface of what is possible with GANs and why learning it would change your thinking about deep neural networks. To know more grab your copy of Kuntal Ganguly’s book on GANs: Learning Generative Adversarial Networks.     .

The world of data science is now starting to change quickly. This was arguably the year when discussions around AI and automation started to escalate, taking on more and more importance in the public sphere. But as interesting as all that is, there are nevertheless real people - like you - actually working with data not to rig elections or steal someone’s jobs but simply to make things better. Arguably, data science and analysis has never been under the spotlight to the extent it is today. Whereas a decade ago there was a whole lotta hope stored in the big data revolution, today there’s anxiety that we’re not doing enough with data, that we don’t have the right data. That makes it a challenging but important time to be working in the world of data. With that in mind, here are our 5 things that will matter in data science in 2018… Find out what 5 data science tools we think will matter most in 2018 here. 1. The ethical considerations in machine learning and artificial intelligence This is huge, but it can’t be ignored. At its heart, this is important because it highlights that there’s human agency at the heart of modern data science, that algorithms are things created and designed by the engineers behind them. But even more important than that, these ethical considerations will be important in 2018 because it will end up defining everyone’s relationship to data for decades to come. And yes, although legislative bodies may play a part in that, it’s also up to people actually working with data to contribute to that discussion about what data does, who uses it and why. That might sound like a lot of responsibility, but it makes things pretty exciting, no? 2. Greater alignment between data projects and business goals This has long been a challenge for just about every business and, indeed, anyone who works in data - architect, analyst or scientist. But as the data hype curve flattens out with more organizations taking advantage of the opportunities it offers, with budgets getting tighter and expectations higher than they have ever been, ensuring that data programs are delivering real value will be crucial in 2018. That means there will be more pressure on data pros to deliver; sharpening your commercial instincts will be essential, and could be your route to the next step in your career. 3. Automated machine learning If budgets are getting tighter and management expectations are higher than ever, the emergence of automated machine learning will be a godsend for 2018. Automated machine learning isn’t a threat to anyone’s job - it’s simply a way of making the steps of algorithm selection and optimization much faster. If you’ve ever lamented the time you’ve spent tweaking an algorithm only for it not to work as you wanted it to, only to move to a further iteration to find a similar problem, automated machine learning will automate away all those iterations. What this means is that you’ll be able to spend more time on value-adding activities that will never be automated away. And in turn this will make you a more valuable data scientist. 4. Taking advantage of cloud Cloud has been a big trend for some years now. But as a word on it’s own it’s always felt a bit abstract and amorphous. However, it’s once you start to see how it can be put into practice that you begin to see how potentially transformative it might be. In the case of machine learning, cloud becomes a vital solution in the battle for resources - it makes machine learning at scale more accessible to more people. The key tool here is Google’s cloud machine learning engine - it’s been built to make building machine learning models as straightforward as possible. When you look at this alongside automated machine learning, it’s possible to suggest that the data science skill set might change somewhat throughout 2018… 5. Better self-service BI 2018 is the year when all employees will need to be empowered by data. The idea that a specific team handles everything relating to data will end; using data will be crucial to a range of different stakeholders. This doesn’t mean the end of the data scientist - as said earlier, no one is going to be losing their jobs. But it does mean that self-service BI tools are going to take on greater importance than ever before in 2018. That means data scientists may have to start thinking more like data architects (especially if there’s no data architect in their organization), and taking into consideration how they make their work accessible and meaningful for stakeholders all around their organization.

Isn’t it spooky how Facebook can identify faces of your friends before you manually tag them? Have you been startled by Cortana, Siri or Google Assistant when they instantly recognize and act on your voice as you speak to these virtual assistants? Deep Learning is the driving force behind these uncanny yet innovative applications. The next thing that is all set to dive into deep learning is the music industry. Neural networks not only ease production and generation of songs, but also assist in music recommendation, transcription and classification. Here are some ways that deep learning will elevate music and the listening experience itself: Generating melodies with Neural Nets At the most basic level, a deep learning algorithm follows 3 simple steps for music generation: First, the neural net is trained with a sample data set of songs which are labelled. The labelling is done based on the emotions you want your song to convey (happy, sad, funny, etc). For training, the program converts the speech of the data set in text format and then creates vector for each word.  The training data can also be in the form of MIDI format which is a standard protocol for encoding musical notes. After completing the training, the program is fed with a set of emotions as input. It identifies the associated input vectors and compares them to training vectors. The output is a melody or chords that represent the desired emotions. Long short-term memory (LSTM) architectures are also used for music generation. They take structured input of a music notation. These inputs are then encoded as vectors and fed into an LSTM at each timestep. LSTM then predicts the encoding of the next timestep. Fully connected convolutional layers are utilized to increase the music quality and to represent rich features in the frequency domain. Magenta, the popular art and music project of Google has launched Performance RNN, which is an LSTM-based recurrent neural network. It is designed to produce multiple sounds with expressive timing and dynamics. In other words, Performance RNN determines which notes to play, when to play them, and how hard to strike each note. IBM’s Watson Beat uses a neural network to produce complete tracks by understanding music theory, structure, and emotional intent. According to Richard Daskas, a music composer working on the Watson Beat project, “Watson only needs about 20 seconds of musical inspiration to create a song.” Transcripting music with deep learning Deep learning methods can also be used for arranging a piece of music for a different instrument. LSTM networks are a popular choice for music transcription and modelling. These networks are trained using a large dataset of pre-labelled music transcriptions (expressed with ABC notation). These transcriptions are then used to generate new music transcriptions. In fact, transformed audio data can be used to predict the group of notes currently being played. This can be achieved by treating the transcription model as an image classification problem. For this, an image of an audio is used, called as Spectrogram. A spectrogram displays how the spectrum or frequency content changes over time. A Short Time Fourier Transform (STFT) or a constant Q transform is used to create this spectrogram. The spectrogram is then feeded to a Convolutional Neural network(CNN). The CNN estimates current notes from audio data and determines what specific notes are present by analysing 88 output nodes for each of the piano keys. This network is generally trained using large number of examples from MIDI files spanning several different genres of music. Magenta has developed The NSynth dataset, which is a high-quality multi-note dataset for music transcription. It is inspired by image recognition datasets and has a huge collection of annotated musical notes. Make better music recommendations Neural Nets are also used to make intelligent music recommendations and are a step ahead of the traditional Collaborative filtering networks. Using neural networks, the system can analyse the songs saved by the users, and then utilize those songs to make new recommendations.  Neural nets can also be used to analyze songs based on musical qualities such as pitch, chord progression, bass, etc. Using the similarities between songs having the same traits as each other, neural networks can detect and predict new songs.  Thus providing recommendation based on similar lyrical and musical styles. Convolutional neural networks (CNNs) are utilized for making music recommendations. A time-frequency representation of the audio signal is fed into the network as the input. 3 second audio clips are randomly chosen from the audio samples to train the neural network. The CNNs are then used to predict latent factors from music audio by taking the average of the predictions for consecutive clips. The feature extraction layers and pooling layers permits operation on several timescales. Spotify is working on a music recommendation system with a CNN. This recommendation system, when trained on short clips of songs, can create playlists based on the audio content only. Classifying music according to genre Classifying music according to a genre is another achievement of neural nets. At the heart of this application lies the LSTM network. At the very first stage, convolutional layers are used for feature extraction from the spectrograms of the audio file. The sequence of features so obtained is given as input to the LSTM layer. LSTM evaluates dependencies of the song across both short time period as well as long term structure. After the LSTM, the input is fed into a fully connected, time-distributed layer which essentially gives us a sequence of vectors. These vectors are then used to output the network's evaluation of the genre of the song at the particular point of time. Deepsound uses GTZAN dataset and an LSTM network to create a model for music genre recognition.  On comparing the mean output distribution with the correct genre, the model gives almost 67% of accuracy. For musical pattern extraction, MFCC feature dataset is used for audio analysis. First, the audio signal is extracted in the MFCC format. Next, the input song is modified into an MFCC map. This Map is then split to feed it as the input of the CNN. Supervised learning is used for automatically obtaining musical pattern extractors, considering the song label is provided. The extractors so acquired, are used for restoring high-order pattern-related features. After high-order classification, the result is combined and undergoes a voting process to produce the song-level label. Scientists from Queen Mary University of London trained a neural net with over 6000 songs in a ballad, hip-hop, and dance to develop a neural network that achieves almost 75% accuracy in song classification. The road ahead Neural networks have advanced the state of music to whole new level where one would no longer require physical instruments or vocals to compose music. The future would see more complex models and data representations to understand the underlying melodic structure. This would help models create compelling artistic content on their own. Combination of music with technology would also foster a collaborative community consisting of artists, coders and deep learning researchers, leading to a tech-driven, yet artistic future.  

While you were away attending NIPS 2017 this week, a lot has been happening around you in the data science and machine learning space. No worries! Here is a brief roundup of the best of what we published on the Datahub this week for your weekend reading. [box type="shadow" align="" class="" width=""]If you would like to share your insights and takeaways from NIPS with our readers on the DataHub, write to us at [email protected].[/box] NIPS 2017 Highlights - Part 1 3 great ways to leverage Structures for Machine Learning problems by Lise Getoor at NIPS 2017 Top Research papers showcased at NIPS 2017 – Part 2 Top Research papers showcased at NIPS 2017 – Part 1 Watch out for more in this area in the coming weeks. Expert in Focus Kate Crawford, Principal Researcher at Microsoft Research and a Distinguished Research Professor at New York University, on 20 lessons on bias in machine learning systems, Keynote at NIPS 2017 3 Things that happened this week in Data Science News PyTorch 0.3.0 releases, ending stochastic functions DeepVariant: Using Artificial Intelligence in Human Genome Sequencing Amazon unveils Sagemaker: An end-to-end machine learning service For a more comprehensive roundup of top news stories this week, check out our weekly news roundup post.   Get hands-on with these Tutorials Understanding Streaming Applications in Spark SQL Implementing Linear Regression Analysis with R What are Slowly Changing Dimensions (SCD) and why you need them in your Data Warehouse? Do you agree with these Insights & Opinions? 4 popular algorithms for Distance-based outlier detection Stitch Fix: Full Stack Data Science and other winning strategies Admiring the many faces of Facial Recognition with Deep Learning One Shot Learning: Solution to your low data problem

The data science tech market is buzzing with new and interesting Machine Learning libraries and tools almost everyday. In an increasingly growing market, it becomes difficult to choose the right tool or set of tools. More importantly, Artificial Intelligence and Deep Learning based projects require a different approach than traditional programming which makes things tricky to zero-in on one library or a framework. The choice of a framework is largely based upon the type of problem, one is expecting to solve. But there are other considerations too. Speed is one such factor that more or less would always play an important role in decision making. Other reasons could be how open-ended it is, architecture, functions, complexity of use, support for algorithms, and so on. Here, we present to you six Java libraries for your next Deep Learning and Artificial Intelligence project you shouldn’t miss if you are a Java loyalist or simply a web developer who wants to enter the world of deep learning. DeepLearning4j (DL4J) One of the first, commercial grade, and most popular deep learning frameworks developed in Java. It also supports other JVM languages (Java, Clojure, Scala). What’s interesting about the DL4J, is that it comes with an in-built GPU support for the training process. It also supports Hadoop YARN for distributed application management. It is popular for solving problems related to image recognition, fraud detection and NLP. MALLET Mallet (Machine Learning for Language Toolkit) is an open source Java Machine Learning toolkit. It supports NLP, clustering, modelling, and classification. The most important capability of Mallet is its support for a wide variety of algorithms such as Naive Bayes and Decision Trees. Another useful feature it has is topic modelling toolkit. Topic models are useful when analyzing large collections of unlabelled texts.   Massive Online Analysis (MOA) MOA is an open source data streaming and mining framework for real time analytics. It has a strong and growing community and is similar and related to Weka. It also has the ability to deal with massive data streams. Encog This framework supports a wide array of algorithms and neural networks such as Artificial Neural Network, Bayesian Network, Genetic Programming and algorithms. Neuroph Neuroph as the name suggests offers great simplicity when working on neural networks. The main USP of Neuroph is its incredibly useful GUI (Graphical User Interface) tool that helps in creating and training neural networks. Neuroph is a good choice of framework when you have a quick project on hand and you don’t want to spend hours learning the theory. Neuroph helps you quickly set up and running in putting neural networks to work for your project. Java Machine Learning Library The Java Machine Learning Library offers a great set of reference implementation of algorithms that you can’t miss for your next Machine Learning project. Some of the key highlights are support vector machines and clustering algorithms. These are a few key frameworks and tools you might want to consider when working on your next research work. The Java ML library ecosystem is vast with many tools and libraries to support, and we just touched the tip of that iceberg in this article. One particular tool that deserve an honourable mention is Environment for Developing KDD-Applications Supported by Index-Structure (ELKI). It is designed particularly with researchers and research students kept in mind. The main focus of ELKI is its broad coverage of data algorithms which makes it a natural fit for research work. What’s really important while choosing any of the above or tools outside of the list is a good understanding of the requirements and the problems you intend to solve. To reiterate, some of the key considerations to bear in mind before zeroing in on a tool would be - support for algorithms, implementation of neural networks, dataset size (small, medium, large), and speed.

Facial recognition technology is not new. In fact, it has been around for more than a decade. However, with the recent rise in artificial intelligence and deep learning, facial technology has achieved new heights. In addition to facial detection, modern day facial recognition technology also recognizes faces with high accuracy and in unfavorable conditions. It can also recognize expressions and analyze faces to generate insights about an individual. Deep learning has enabled a power-packed face recognition system, all geared up to achieve widespread adoption. How has deep learning modernised facial recognition Traditional facial recognition algorithms would recognize images and people using distinct facial features (placement of eye, eye color, nose shape etc.) However, they failed in correct identification in cases of different lighting or slight change in the appearance ( beard growth, aging, or pose). In order to develop facial recognition techniques for a dynamic and ever-changing face, deep learning is proving to be a game changer. Deep Neural nets go beyond the approach of manual extraction. These AI based Neural Networks rely on image pixels to analyze features of a particular face. So they scan faces irrespective of the lighting, ageing, pose, or emotions. Deep learning algorithms remember each time they recognize or fail to recognize a problem. Thus, avoiding repeat mistakes and getting better at each attempt. Deep learning algorithms can also be helpful in converting 2D images to 3D. Facial recognition in practice: Facial Recognition Technology in Multimedia Deep learning enabled facial recognition technologies can be used to track audience reaction and measure different levels of emotions. Essentially it can predict how a member of the audience will react to the remaining film. Not only this, it also helps determine what percentage of users will be interested in a particular movie genre. For example, Microsoft’s Azure Emotion,  an emotion API detects emotions by analysing the facial expressions on an image or video content over time. Caltech and Disney have collaborated to develop a neural network which can track facial expressions. Their deep learning based Factorised Variational Autoencoders (FVAEs) analyze facial expressions of audience for about 10 minutes and then predict how their reaction will be for the rest of the film. These techniques help in estimating whether the viewers are giving the expected reactions at the right place. For example, the viewer is not expected to yawn on a comical scene. With this, Disney can also predict the earning potential of a particular movie. It can generate insights that may help producers create compelling movie trailers to maximize the number of footfalls. Smart TVs are also equipped with sophisticated cameras and deep learning algos for facial recognition ability. They can recognize the face of the person watching and automatically show channels and web applications programmed as their favorites. The British broadcasting corporation uses the facial recognition technology, built by CrowdEmotion. By tracking faces of almost 4,500 audience members watching show trailers, they gauge exact customer emotions about a particular programme. This in turn helps them generate insights to showcase successful commercials. Biometrics in Smartphones A large number of smartphones nowadays are instilled with biometric capabilities. Facial recognition in smartphones are not only used as a means of unlocking and authorizing, but also for making secure transactions and payments. In present times, there has been a rise in chips with built-in deep learning ability. These chips are embedded into smartphones. By having a neural net embedded inside the device, crucial face biometric data never leaves the device or sent to the cloud. This in turn improves privacy and reduces latency. Some of the real-world examples include Intel’s Nervana Neural Network Processor, Google’s TPU, Microsoft’s FPGA, and Nvidia’s Tesla V100. Deep learning models, embedded in a smartphone, can construct a mathematical model of the face which is then stored in the database. Using this mathematical face model, smartphones can easily recognize users even as their face ages or when it is obstructed by wearable accessories. Apple has recently launched the iPhone X facial recognition system termed as FaceID. It maps thousands of points on a user’s face using a projector and an infrared camera (which can operate under varied lighting conditions). This map is then passed to a bionic chip embedded in the smart phone. The chip has a neural network which constructs a mathematical model of the user’s face, used for biometric face verification and recognition. Windows Hello is also a facial recognition technology to unlock Windows smart devices equipped with infrared cameras. Qualcomm, a mobile technology organization, is working on a new depth-perception technology. It will include an image signal processor and high-resolution 3D depth-sensing cameras for facial recognition. Face recognition for Travel Facial recognition technologies can smoothen the departure process for a customer by eliminating the need for a boarding pass. A traveller is scanned by cameras installed at various check points, so they don’t have to produce a boarding pass at every step. Emirates is collaborating with Dubai Customs, Police and Airports to use a facial recognition technology solution integrated with the UAE Wallet app. The project is known as Together Initiative, it allows travellers to register and store their biometric facial data at several kiosks placed at the check-in area. This facility helps passengers to avoid presenting their physical documents at every touchpoint. Face recognition can also be used for determining illegal immigration. The technology compares the photos of passengers taken immediately before boarding, with the photos provided in their visa application. Biometric Exit, is an initiative by US government, which uses facial recognition to identify individuals leaving the country. Facial recognition technology can also be used at train stations to reduce the waiting time for  buying a train ticket or going through other security barriers. Bristol Robotics Laboratory has developed a software which uses infrared cameras to identify passengers as they walk onto the train platform. They do not need to carry tickets. Retail and shopping In the area of retail, smart facial recognition technologies can be helpful in fast checkout by keeping a track of each customer as they shop across a store. This smart technology, can also use machine learning and analytics to find trends in the shopper’s purchasing behavior over time and devise personalized recommendations. Facial video analytics and deep learning algorithms can also identify loyal and VIP shoppers from the moving crowd, giving them a privileged VIP experience. Thus, enabling them with more reasons to come back and make repeat purchases. Facial biometrics can also accumulate rich statistics about demographics(age, gender, shopping history) of an individual. Analyzing these statistics can generate insights, which helps organizations develop their products and marketing strategies. FindFace is one such platform that uses sophisticated deep learning technologies to generate meaningful data about the shopper. Its e-facial recognition system can verify faces with almost 99% accuracy. It can also help route the shopper data to a salesperson’s notice for personalized assistance. Facial recognition technology can also be used to make secure payment transactions simply by analysing a person’s face. AliBaba has set up a Smile to Pay face recognition system in KFC's. This system allows customers to make secure payments by merely scanning their face. Facial recognition has emerged as a hot topic of interest and is poised to grow. On the flip side, organizations deploying such technology should incorporate privacy policies as a standard measure. Data collected from such facial recognition software can also be used wrongly for targeting customers with ads, or for other illegal purposes. They should implement a methodical and systematic approach for using facial recognition for the benefit of their customers. This will not only help businesses generate a new source of revenue, but will also usher in a new era of judicial automation.  

Last week, a company in San Francisco was popping bottles of champagne for their achievements. And trust me, they’re not at all small. Not even a couple of weeks gone by, since it was listed on the stock market and it has soared to over 50%. Stitch Fix is an apparel company run by co-founder and CEO, Katrina Lake. In just a span of 6 years, she’s been able to build the company with an annual revenue of a whopping $977 odd million. The company has been disrupting traditional retail and aims to bridge the gap of personalised shopping, that the former can’t accomplish. Stitch Fix is more of a personalized stylist, rather than a traditional apparel company. It works in 3 basic steps: Filling a Style Profile: Clients are prompted to fill out a style profile, where they share their style, price and size preferences. Setting a Delivery Date: The clients set a delivery date as per their availability. Stitch Fix mixes and matches various clothes from their warehouses and comes up with the top 5 clothes that they feel would best suit the clients, based on the initial style profile, as well as years of experience in styling. Keep or Send Back: The clothes reach the customer on the selected date and the customer can try on the clothes, keep whatever they like or send back what they don’t. The aim of Stitch Fix is to bring a personal touch to clothes shopping. According to Lake, “There are millions and millions of products out there. You can look at eBay and Amazon. You can look at every product on the planet, but trying to figure out which one is best for you is really the challenge” and that’s the tear Stitch Fix aims to sew up. In an interview with eMarketer, Julie Bornstein, COO of Stitch Fix said “Over a third of our customers now spend more than half of their apparel wallet share with Stitch Fix. They are replacing their former shopping habits with our service.” So what makes Stitch Fix stand out among its competitors? How do they do it? You see, Stitch Fix is not just any apparel company. It has created the perfect formula by blending human expertise with just the right amount of Data Science to enable it to serve its customers. When we’re talking about the kind of Data Science that Stitch Fix does, we’re talking about a relatively new and exciting term that’s on the rise - Full Stack Data Science. Hello Full Stack Data Science! For those of you who’ve heard of this before, cheers! I hope you’ve had the opportunity to experience its benefits. For those of you who haven’t heard of the term, Full Stack Data Science basically means a single data scientist does their own work, which is mining data, cleans it, writes an algorithm to model it and then visualizes the results, while also stepping into the shoes of an engineer, implementing the model, as well as a Project Manager, tracking the entire process and ensuring it’s on track. Now while this might sound like a lot for one person to do, it’s quite possible and practical. It’s practical because of the fact that when these roles are performed by different individuals, they induce a lot of latency into the project. Moreover, a synchronization of priorities of each individual is close to impossible, thus creating differences within the team. The Data (Science) team at Stitch Fix is broadly categorized based on what area they work on: Because most of the team focuses on full stack, there are over 80 Data Scientists on board. That’s a lot of smart people in one company! On a serious note, although unique, this kind of team structure has been doing well for them, mainly because it gives each one the freedom to work independently. Tech Treasure Trove When you open up Stitch Fix’s tech toolbox, you won’t find Aladdin’s lamp glowing before you. Their magic lies in having a simple tech stack that works wonders when implemented the right way. They work with Ruby on Rails and Bootstrap for their web applications that are hosted on Heroku. Their data platform relies on a robust Postgres implementation. Among programming languages, we found Python, Go, Java and JavaScript also being used. For an ML Framework, we’re pretty sure they’re playing with TensorFlow. But just working with these tools isn’t enough to get to the level they’re at. There’s something more under the hood. And believe it or not, it’s not some gigantic artificial intelligent system running on a zillion cores! Rather, it’s all about the smaller, simpler things in life. For example, if you have 3 different kinds of data and you need to find a relationship between them, instead of bringing in the big guns (read deep learning frameworks), a simple tensor decomposition using word vectors would do the deed quite well. Advantages galore: Food for the algorithms One of the main advantages Stitch Fix has, is that they have almost 5 years’ worth client data. This data is obtained from clients in several ways like through a Client Profile, After-Delivery Feedback, Pinterest photos, etc. All this data is put through algorithms that learn more about the likes and dislikes of clients. Some interesting algorithms that feed on this sumptuous data are on the likes of collaborative filtering recommenders to group clients based on their likes, mixed-effects modeling to learn about a client’s interests over time, neural networks to derive vector descriptions of the Pinterest images and to compare them with in-house designs, NLP to process customer feedback, Markov chain models to predict demand, among several others. A human Touch: When science meets art While the machines do all the calculations and come up with recommendations on what designs customers would appreciate, they still lack the human touch involved. Stitch Fix employs over 3000 stylists. Each client is assigned a stylist who knows the entire preference of the client at the glance of a custom-built interface. The stylist finalizes the selections from the inventory list also adding in a personal note that describes how the client can accessorize the purchased items for a particular occasion and how they can pair them with any other piece of clothing in their closet. This truly advocates “Humans are much better with the machines, and the machines are much better with the humans”. Cool, ain't it? Data Platform Apart from the Heroku platform, Stitch Fix seems to have internal SaaS platforms where the data scientists effectively carry out analysis, write algorithms and put them into production. The platforms exhibit properties like data distribution, parallelization, auto-scaling, failover, etc. This lets the data scientists focus on the science aspect while still enjoying the benefits of a scalable system. The good, the bad and the ugly: Microservices, Monoliths and Scalability Scalability is one of the most important aspects a new company needs to take into account before taking the plunge. Using a microservice architecture helps with this, by allowing small independent services/mini applications to run on their own. Stitch Fix uses this architecture to improve scalability although, their database is a monolith. They now are breaking the monolith database into microservices. This is a takeaway for all entrepreneurs just starting out with their app. Data Driven Applications Data-driven applications ensure that the right solutions are built for customers. If you’re a customer-centric organisation, there’s something you can learn from Stitch Fix. Data-Driven Apps seamlessly combine the operational and analytic capabilities of the organisation, thus breaking down the traditional silos. TDD + CD = DevOps Simplified Both Test Driven Development and Continuous Delivery go hand in hand and it’s always better to imbibe this culture right from the very start. In the end, it’s really great to see such creative and technologically driven start-ups succeed and sail to the top. If you’re on the journey to building that dream startup of yours and you need resources for your team, here’s a few books you’ll want to pick up to get started with: Hands-On Data Science and Python Machine Learning by Frank Kane Data Science Algorithms in a Week by Dávid Natingga Continuous Delivery and DevOps : A Quickstart Guide - Second Edition by Paul Swartout Practical DevOps by Joakim Verona    

In 2003, when The New York Times announced that the human genome project was successfully complete two years ahead of its schedule (leave aside the conspiracy theory that the genome was never ‘completely’ sequenced), it heralded a new dawn in the history of modern science. The challenge thereafter was to make sense out of the staggering data that became available. The High Throughput Sequencing technology came to revolutionize the processing of genomic data in a way, but had its own limitations (such as the high rate of erroneous base calls produced). Google has now launched an artificial intelligence tool, DeepVariant, to analyze the huge data resulting from the sequencing of the genome. It took two years of research for Google to build DeepVariant. It's a combined effort from Google’s Brain team, a group that focuses on developing and applying AI techniques, and Verily Life Sciences, another Alphabet subsidiary that is focused on the life sciences. How the DeepVariant makes sense of your genome? DeepVariant uses the latest deep learning techniques to turn high-throughput sequencing readouts into a picture of a full genome. It automatically identifies small insertion and deletion mutations and single-base-pair mutations in sequencing data. Ever since the high-throughput sequencing made genome sequencing more accessible, the data produced has at best offered error-prone snapshot of a full genome. Researchers have found it challenging to distinguish small mutations from random errors generated during the sequencing process, especially in repetitive portions of a genome. A number of tools and methods have come out to interpret these readouts (both public and private funded), but all of them have used simpler statistical and machine-learning approaches to identify mutations. Google claims DeepVariant offers significantly greater accuracy than all previous classical methods. DeepVariant transforms the task of variant calling (the process to identify variants from sequence data) into an image classification problem well-suited to Google's existing technology and expertise. Google's team collected millions of high-throughput reads and fully sequenced genomes from the Genome in a Bottle (GIAB) project, and fed the data to a deep-learning system that interpreted sequenced data with a high level of accuracy. “Using multiple replicates of GIAB reference genomes, we produced tens of millions of training examples in the form of multi-channel tensors encoding the HTS instrument data, and then trained a TensorFlow-based image classification model to identify the true genome sequence from the experimental data produced by the instruments.” Google said. The result has been remarkable. Within a year, DeepVariant went on to win first place in the PrecisionFDA Truth Challenge, outperforming all state-of-the-art methods in accurate genetic sequencing. “Since then, we've further reduced the error rate by more than 50%,” the team claims. Image Source: research.googleblog.com “The success of DeepVariant is important because it demonstrates that in genomics, deep learning can be used to automatically train systems that perform better than complicated hand-engineered systems,” says Brendan Frey, CEO of Deep Genomics, one of the several companies using AI on genomics for potential drugs. DeepVariant is ‘open’ for all The best thing about DeepVariant is that it has been launched as an open source software. This will encourage enthusiastic researchers for collaboration and possibly accelerate its adoption to solve real world problems. “To further this goal, we partnered with Google Cloud Platform (GCP) to deploy DeepVariant workflows on GCP, available today, in configurations optimized for low-cost and fast turnarounds using scalable GCP technologies like the Pipelines API,” Google said. This paired set of releases could facilitate a scalable, cloud-based solution to handle even the largest genomics datasets. The road ahead: What DeepVariant means for future According to Google, DeepVariant is the first of “what we hope will be many contributions that leverage Google's computing infrastructure and Machine learning expertise” to better understand the genome and provide deep learning-based genomics tools to the community. This is, in fact, all part of a “broader goal” to apply Google technologies to healthcare and other scientific applications. As AI starts to propel different branches of medicine take big leaps forward in coming years, there is a whole lot of medical data to mine and drive insights from. But with genomic medicine, the scale is huge. We are talking about an unprecedented set of data that is equally complex. “For the first time in history, our ability to measure our biology, and even to act on it, has far surpassed our ability to understand it,” says Frey. “The only technology we have for interpreting and acting on these vast amounts of data is AI. That’s going to completely change the future of medicine.” These are exciting times for medical research. In 1990, when the human genome project was initiated, it met with a lot of skepticism from many people, including scientists and non-scientists alike. But today, we have completely worked out each A, T, C, and G that makes up the DNA of all 23 pairs of human chromosomes. After high-throughput sequencing made the genomic data accessible, Google’s DeepVariant could just be the next big thing to take genetic sequencing to a whole new level.

The fact that machines are successful in replicating human intelligence is mind-boggling. However, this is only possible if machines are fed with correct mix of algorithms, huge collection of data, and most importantly the training given to it, which in turn leads to faster prediction or recognition of objects within the images. On the other hand, when you train humans to recognize a car for example, you simply have to show them a live car or an image. The next time they see any vehicle, it would be easy for them to distinguish a car among-st other vehicles. In a similarly way, can machines learn with single training example like humans do? Computers or machines lack a key part that distinguishes them from humans, and that is, ‘Memory’. Machines cannot remember; hence it requires millions of data to be fed in order to understand the object detection, be it from any angle. In order to reduce this supplement of training data and enabling machines to learn with less data at hand, One shot learning is brought to its assistance. What is one shot learning and how is it different from other learning? Deep Neural network models outperform various tasks such as image recognition, speech recognition and so on. However, such tasks are possible only due to extensive, incremental training on large data sets. In cases when there is a smaller dataset or fewer training examples, a traditional model is trained on the data that is available. During this process, it relearns new parameters and incorporates new information, and completely forgets the one previously learned. This leads to poor training or catastrophic inference. One shot learning proves to be a solution here, as it is capable of learning with one, or a minimal number of training samples, without forgetting. The reason for this is, they posses meta-learning; a capability often seen in neural network that has memory. How One shot learning works? One shot learning strengthens the ability of the deep learning models without the need of a huge dataset to train on. Implementation of One shot learning can be seen in a Memory Augmented Neural Network (MANN) model. A MANN has two parts, a controller and an external memory model. The controller is either a feed forward neural network or an LSTM (Long Short Term Memory) network, which interacts with the external memory module using number of read/write heads. These heads fetch or place representations to and fro the memory. LSTMs are proficient in long term storage through slow updates of weights and short term storage via the external memory module. They are trained to meta-learn; i.e. it can rapidly learn unseen functions with fewer data samples.Thus, MANNs are said to be capable of metalearning. The MANN model is later trained on datasets that include different classes with very few samples. For instance, the Omniglot dataset, a collection of handwritten samples of different languages, with very few samples of each language. After continuously training the model with thousands of iterations by using few samples, the model was able to recognize never-seen-before image samples, taken from a disjoint sample of the Omniglot dataset. This proves that MANN models are able to outperform various object categorization tasks with minimal data samples.   Similarly, One shot learning can also be achieved using Neural Turing Machine and Active One shot learning. Therefore, learning with a single attempt/one shot actually involves meta-learning. This means, the model gradually learns useful representations from the raw data using certain algorithms, for instance, the gradient descent algorithm. Using these learnings as a base knowledge, the model can rapidly cohere never seen before information with a single or one-shot appearance via an external memory module. Use cases of One shot learning Image Recognition: Image representations are learnt using supervised metric based approach. For instance,  siamese neural network, an identical sister network, discriminates between the class-identity of an image pair. Features of this network are reused for one-shot learning without the need for retraining. Object Recognition within images: One shot learning allows neural network models to recognize known objects and its category within an image. For this, the model learns to recognize the object with a few set of training samples. Later it compares the probability of the object to be present within the image provided. Such a model trained on one shot can recognize objects in an image despite the clutter, viewpoint, and lighting changes.   Predicting accurate drugs: The availability of datasets for a drug discovery are either limited or expensive. The molecule found during a biological study often does not end up being a drug due to ethical reasons such as toxicity, low-solubility and so on. Hence, a less amount of data is available about the candidate molecule. Using one shot learning, an iterative LSTM combined with Graph convolutional neural network is used to optimize the candidate molecule. This is done by finding similar molecules with increased pharmaceutical activity and lesser risks to patients. A detailed explanation of how using low data, accurate drugs can be predicted is discussed in a research paper published by the American Chemical Society(ACS). One shot learning is in its infancy and therefore use cases can be seen in familiar applications such as image and object recognition. As the technique will advance with time and the rate of adoption, other applications of one shot learning will come into picture. Conclusion One shot learning is being applied in instances of machine learning or deep learning models that have less data available for their training. A plus point in future is, that organizations will not have to collect huge amount of data for their ML models to be trained, only a few training samples would do the job! Large number of organizations are looking forward to adopt one shot learning within their deep learning models. It would be exciting to see how one shot learning will glide through being the base of every neural network implementation.  

Imagine yourself playing a strategy game, like Age of Empires perhaps. You are in a world that looks real and you are pitted against the computer, and your mission is to protect your empire and defeat the computer, at the same time. What if you could create an army of soldiers who could explore the map and attack the enemies on their own, based on just a simple command you give them? And what if your soldiers could have real, unscripted conversations with you as their commander-in-chief to seek instructions? And what if the game’s scenes change spontaneously based on your decisions and interactions with the game elements, like a movie? Sounds too good to be true? It’s not far-fetched at all - thanks to the rise of Artificial Intelligence! The gaming industry today is a market worth over a hundred billion dollars. The Global Games Market Report says that about 2.2 billion gamers across the world are expected to generate an incredible $108.9 billion in game revenue by the end of 2017. As such, gaming industry giants are seeking newer and more innovative ways to attract more customers and expand their brands. While terms like Virtual Reality, Augmented Reality and Mixed Reality come to mind immediately as the future of games, the rise of Artificial Intelligence is an equally important stepping stone in making games smarter and more interactive, and as close to reality as possible. In this article, we look at the 5 ways AI is revolutionizing the gaming industry, in a big way! Making games smarter While scripting is still commonly used for control of NPCs (Non-playable character) in many games today, many heuristic algorithms and game AIs are also being incorporated for controlling these NPCs. Not just that, the characters also learn from the actions taken by the player and modify their behaviour accordingly. This concept can be seen implemented in Nintendogs, a real-time pet simulation video game by Nintendo. The ultimate aim of the game creators in the future will be to design robust systems within games that understand speech, noise and other sounds within the game and tweak the game scenario accordingly. This will also require modern AI techniques such as pattern recognition and reinforcement learning, where the characters within the games will self-learn from their own actions and evolve accordingly. The game industry has identified this and some have started implementing these ideas - games like F.E.A.R and The Sims are a testament to this. Although the adoption of popular AI techniques in gaming is still quite limited, their possible applications in the near-future has the entire gaming industry buzzing. Making games more realistic This is one area where the game industry has grown leaps and bounds over the last 10 years. There have been incredible advancements in 3D visualization techniques, physics-based simulations and more recently, inclusion of Virtual Reality and Augmented Reality in games. These tools have empowered game developers to create interactive, visually appealing games which one could never imagine a decade ago. Meanwhile, gamers have evolved too. They don’t just want good graphics anymore; they want games to resemble reality. This is a massive challenge for game developers, and AI is playing a huge role in addressing this need. Imagine a game which can interpret and respond to your in-game actions, anticipate your next move and act accordingly. Not the usual scripts where an action X will give a response Y, but an AI program that chooses the best possible alternative to your action in real-time, making the game more realistic and enjoyable for you. Improving the overall gaming experience Let’s take a real-world example here. If you’ve played EA Sports’ FIFA 17, you may be well-versed with their Ultimate Team mode. For the uninitiated, it’s more of a fantasy draft, where you can pick one of the five player choices given to you for each position in your team, and the AI automatically determines the team chemistry based on your choices. The team chemistry here is important, because the higher the team chemistry, the better the chances of your team playing well. The in-game AI also makes the playing experience better by making it more interactive. Suppose you’re losing a match against an opponent - the AI reacts by boosting your team’s morale through increased fan chants, which in turn affects player performances positively. Gamers these days pay a lot of attention to detail - this not only includes the visual appearance and the high-end graphics, but also how immersive and interactive the game is, in all possible ways. Through real-time customization of scenarios, AI has the capability to play a crucial role in taking the gaming experience to the next level. Transforming developer skills The game developer community have always been innovators in adopting cutting edge technology to hone their technical skills and creativity. Reinforcement Learning, a sub-set of Machine Learning, and the algorithm behind the popular AI computer program AlphaGo, that beat the world’s best human Go player is a case in point. Even for the traditional game developers, the rising adoption of AI in games will mean a change in the way games are developed. In an interview with Gamasutra, AiGameDev.com’s Alex Champandard says something interesting: “Game design that hinges on more advanced AI techniques is slowly but surely becoming more commonplace. Developers are more willing to let go and embrace more complex systems.” It’s safe to say that the notion of Game AI is changing drastically. Concepts such as smarter function-based movements, pathfinding, inclusion of genetic algorithms and rule-based AI such as fuzzy logic are being increasingly incorporated in games, although not at a very large scale. There are some implementation challenges currently as to how academic AI techniques can be brought more into games, but with time these AI algorithms and techniques are expected to embed more seamlessly with traditional game development skills. As such, in addition to knowledge of traditional game development tools and techniques, game developers will now have to also skill up on these AI techniques to make smarter, more realistic and more interactive games. Making smarter mobile games The rise of the mobile game industry today is evident from the fact that close to 50% of the game revenue in 2017 will come from mobile games - be it smartphones or tablets. The increasingly high processing power of these devices has allowed developers to create more interactive and immersive mobile games. However, it is important to note that the processing power of the mobile games is yet to catch up to their desktop counterparts, not to mention the lack of a gaming console, which is beyond comparison at this stage. To tackle this issue, mobile game developers are experimenting with different machine learning and AI algorithms to impart ‘smartness’ to mobile games, while still adhering to the processing power limits. Compare today’s mobile games to the ones 5 years back, and you’ll notice a tremendous shift in terms of the visual appearance of the games, and how interactive they have become. New machine learning and deep learning frameworks & libraries are being developed to cater specifically to the mobile platform. Google’s TensorFlow Lite and Facebook’s Caffe2 are instances of such development. Soon, these tools will come to developers’ rescue to build smarter and more interactive mobile games. In Conclusion Gone are the days when games were just about entertainment and passing time. The gaming industry is now one of the most profitable industries of today. As it continues to grow, the demands of the gaming community and the games themselves keep evolving. The need for realism in games is higher than ever, and AI has an important role to play in making games more interactive, immersive and intelligent. With the rate at which new AI techniques and algorithms are developing, it’s an exciting time for game developers to showcase their full potential. Are you ready to start building AI for your own games? Here are some books to help you get started: Practical Game AI Programming Learning game AI programming with Lua

Learning something by rote i.e., repeating it many times, perfecting a skill by practising it over and over again or building something by making minor adjustments progressively to a prototype are things that comes to us naturally as human beings. Machines can also learn this way and this is called ‘Iterative machine learning’. In most cases, iteration is an efficient learning approach that helps reach the desired end results faster and accurately without becoming a resource crunch nightmare. Now, you might wonder, isn’t iteration inherently part of any kind of machine learning? In other words, modern day machine learning techniques across the spectrum from basic regression analysis, decision trees, Bayesian networks, to advanced neural nets and deep learning algorithms have some inherent iterative component built into them. What is the need, then, for discussing iterative learning as a standalone topic? This is simply because introducing iteration externally to an algorithm can minimize the error margin and therefore help in accurate modelling.  How Iterative Learning works Let’s understand how iteration works by looking closely at what happens during a single iteration flow within a machine learning algorithm. A pre-processed training dataset is first introduced into the model. After processing and model building with the given data, the model is tested, and then the results are matched with the desired result/expected output. The feedback is then returned back to the system for the algorithm to further learn and fine tune its results. This clearly shows that two iteration processes take place here: Data Iteration - Inherent to the algorithm Model Training Iteration - Introduced externally Now, what if we did not feedback the results into the system i.e. did not allow the algorithm to learn iteratively but instead adopted a sequential approach? Would the algorithm work and would it provide the right results? Yes, the algorithm would definitely work. However, the quality of the results it produces is going to vary vastly based on a number of factors. The quality and quantity of the training dataset, the feature definition and extraction techniques employed, the robustness of the algorithm itself are among many other factors. Even if all of the above were done perfectly, there is still no guarantee that the results produced by a sequential approach will be highly accurate. In short, the results will neither be accurate nor reproducible. Iterative learning thus allows algorithms to improve model accuracy. Certain algorithms have iteration central to their design and can be scaled as per the data size. These algorithms are at the forefront of machine learning implementations because of their ability to perform faster and better. In the following sections we will discuss iteration in different sets of algorithms each from the three main machine learning approaches - supervised ML, unsupervised ML and reinforcement learning. The Boosting algorithms: Iteration in supervised ML The boosting algorithms, inherently iterative in nature, are a brilliant way to improve results by minimizing errors. They are primarily designed to reduce bias in results and transform a particular set of weak learning classifier algorithms to strong learners and to enable them to reduce errors. Some examples are: AdaBoost (Adaptive Boosting) Gradient Tree Boosting XGBoost How they work All boosting algorithms have a common classifiers which are iteratively modified to reach the desired result. Let’s take the example of finding cases of plagiarism in a certain article. The first classifier here would be to find a group of words that appear somewhere else or in another article which would result in a red flag. If we create 10 separate group of words and term them as classifiers 1 to 10, then our article will be checked on the basis of this classifier and any possible matches will be red flagged. But no red flags with these 10 classifiers would not mean a definite 100% original article. Thus, we would need to update the classifiers, create shorter groups perhaps based on the first pass and improve the accuracy with which the classifiers can find similarity with other articles. This iteration process in Boosting algorithms eventually leads us to a fairly high rate of accuracy. The reason being after each iteration, the classifiers are updated based on their performance. The ones which have close similarity with other content are updated and tweaked so that we can get a better match. This process of improving the algorithm inherently, is termed as boosting and is currently one of the most popular methods in Supervised Machine Learning. Strengths & weaknesses The obvious advantage of this approach is that it allows minimal errors in the final model as the iteration enables the model to correct itself every time there is an error. The downside is the higher processing time and the overall memory requirement for a large number of iterations. Another important aspect is that the error fed back to train the model is done externally, which means the supervisor has control over the model and how it modifies. This in turn has a downside that the model doesn’t learn to eliminate error on its own. Hence, the model is not reusable with another set of data. In other words, the model does not learn how to become error-free by itself and hence cannot be ported to another dataset as it would need to start the learning process from scratch. Artificial Neural Networks: Iteration in unsupervised ML Neural Networks have become the poster child for unsupervised machine learning because of their accuracy in predicting data models. Some well known neural networks are: Convolutional Neural Networks   Boltzmann Machines Recurrent Neural Networks Deep Neural Networks Memory Networks How they work Artificial neural networks are highly accurate in simulating data models mainly because of their iterative process of learning. But this process is different from the one we explored earlier for Boosting algorithms. Here the process is seamless and natural and in a way it paves the way for reinforcement learning in AI systems. Neural Networks consist of electronic networks simulating the way the human brain is works. Every network has an input and output node and in-between hidden layers that consist of algorithms. The input node is given the initial data set to perform a set of actions and each iteration creates a result that is output as a string of data. This output is then matched with the actual result dataset and the error is then fed back to the input node. This error then enables the algorithms to correct themselves and reach closer and closer to the actual dataset. This process is called training the Neural Networks and each iteration improve the accuracy. The key difference between the iteration performed here as compared to how it is performed by Boosting algorithms is that here we don’t have to update the classifiers manually, the algorithms change themselves based on the error feedback. Strengths & weaknesses The main advantage of this process is obviously the level of accuracy that it can achieve on its own. The model is also reusable because it learns the means to achieve accuracy and not just gives you a direct result. The flip side of this approach is that the models can go wrong heavily and deviate completely in a different direction. This is because the induced iteration takes its own course and doesn’t need human supervision. The facebook chat-bots deviating from their original goal and communicating within themselves in a language of their own is a case in point. But as is the saying, smart things come with their own baggage. It’s a risk we would have to be ready to tackle if we want to create more accurate models and smarter systems.    Reinforcement Learning Reinforcement learning is a interesting case of machine learning where the simple neural networks are connected and together they interact with the environment to learn from their mistakes and rewards. The iteration introduced here happens in a complex way. The iteration happens in the form of reward or punishment for arriving at the correct or wrong results respectively. After each interaction of this kind, the multilayered neural networks incorporate the feedback, and then recreate the models for better accuracy. The typical type of reward and punishment method somewhat puts it in a space where it is neither supervised nor unsupervised, but exhibits traits of both and also has the added advantage of producing more accurate results. The con here is that the models are complex by design. Multilayered neural networks are difficult to handle in case of multiple iterations because each layer might respond differently to a certain reward or punishment. As such it may create inner conflict that might lead to a stalled system - one that can’t decide which direction to move next. Some Practical Implementations of Iteration Many modern day machine learning platforms and frameworks have implemented the iteration process on their own to create better data models, Apache Spark and MapR are two such examples. The way the two implement iteration is technically different and they have their merits and limitations. Let’s look at MapReduce. It reads and writes data directly onto HDFS filesystem present on the disk. Note that for every iteration to be read and written from the disk needs significant time. This in a way creates a more robust and fault tolerant system but compromises on the speed. On the other hand, Apache Spark stores the data in memory (Resilient Distributed DataSet) i.e. in the RAM. As a result, each iteration takes much less time which enables Spark to perform lightning fast data processing. But the primary problem with the Spark way of doing iteration is that dynamic memory or RAM is much less reliable than disk memory to store iteration data and perform complex operations. Hence it’s much less fault tolerant that MapR.   Bringing it together To sum up the discussion, we can look at the process of iteration and its stages in implementing machine learning models roughly as follows: Parameter Iteration: This is the first and inherent stage of iteration for any algorithm. The parameters involved in a certain algorithm are run multiple times and the best fitting parameters for the model are finalized in this process. Data Iteration: Once the model parameters are finalized, the data is put into the system and the model is simulated. Multiple sets of data are put into the system to check the parameters’ effectiveness in bringing out the desired result. Hence, if data iteration stage suggests that some of the parameters are not well suited for the model, then they are taken back to the parameter iteration stage and parameters are added or modified. Model Iteration: After the initial parameters and data sets are finalized, the model testing/ training happens. The iteration in model testing phase is all about running the same model simulation multiple times with the same parameters and data set, and then checking the amount of error, if the error varies significantly in every iteration, then there is something wrong with either the data or the parameter or both. Iterations are done to data and parameters until the model achieves accuracy.   Human Iteration: This step involves the human induced iteration where different models are put together to create a fully functional smart system. Here, multiple levels of fitting and refitting happens to achieve a coherent overall goal such as creating a driverless car system or a fully functional AI. Iteration is pivotal to creating smarter AI systems in the near future. The enormous memory requirements for performing multiple iterations on complex data sets continue to pose major challenges. But with increasingly better AI chips, storage options and data transfer techniques, these challenges are getting easier to handle. We believe iterative machine learning techniques will continue to lead the transformation of the AI landscape in the near future.