Tech News - Data

At NeurIPS 2017, a group of Stanford and Google researchers presented a very intriguing study on how a neural network, CycleGAN learns to cheat. The researchers trained CycleGAN to transform aerial images into street maps, and vice versa. They found that the neural network learned to hide information about the original image inside the generated one in the form of a low-amplitude high-frequency signal, which almost appears to be noise. Using this information, the generator can then reproduce the original image and thus satisfy the cyclic consistency requirement. What is CycleGAN? CycleGAN is an algorithm for performing image-to-image translation where the neural network needs to learn the mapping between an input image and an output image with the help of a training set of aligned image pairs. What sets CycleGAN apart from other GAN algorithms is that it does not require paired training data. It translates images from a source domain X to a target domain Y without needing paired examples. How CycleGAN was hiding information? For the experiment, the researchers trained CycleGAN on a maps dataset that consisted of 1,000 aerial photographs X and 1,000 maps Y. After training for 500 epochs, the model produced two maps F : X → Y and G : Y → X that generated realistic samples from these image domains. While training the model, the researchers used an aerial photograph that was unseen by the network. The generated image was nearly identical to the source images. On closer inspection, the researchers observed that there are various details present in both the original aerial photograph and the aerial reconstruction that are not visible in the intermediate map, as shown in the following figure: Source: CycleGAN The network showed this result with nearly every aerial photograph, even when it was trained on datasets other than maps. After making this observation, the researchers concluded that CycleGAN is learning an encoding scheme in which it hides information about the aerial photograph within the generated map. What are the implications of this information hiding? The researchers highlighted that this property of encoding information can make this model vulnerable to adversarial attacks. The study showed that CycleGAN can reconstruct any aerial image from a specifically crafted map by starting gradient descent from an initial source map. Attackers can misuse this fact and cause one of the learned transformations to produce an image of their choice by perturbing any chosen source image. Also, if the developers are not careful and are not taking proper measures, their models may be collecting personal data under GDPR. How can these the attacks be avoided? The vulnerability of this model is caused by two reasons: cyclic consistency loss and the difference in entropy between two domains. The cyclic consistency loss can be modified to prevent such attacks. The entropy of one of the domains can be increased artificially by adding an additional hidden variable. The paper, ‘CycleGAN, a Master of Steganography’ grabbed attention when it was posted on Reddit and sparked a discussion. Many Redditors suggested a solution to this, and one of them said, “Adding nearly imperceptible gaussian noise between the cycles should be enough to prevent the CycleGAN from hiding information encoded in imperceptible high-frequency components: it forces it to encode all semantic information in whatever is able to survive low-amplitude gaussian noise (i.e. the visible low-frequency components, as we want/expect).” A recent work towards reducing this steganographic behavior is the introduction of Augmented CycleGAN, which learns many-to-many mappings between domains unlike CycleGAN which learns one-to-one mappings. To know more about this in detail, check out the paper: CycleGAN, a Master of Steganography. What are generative adversarial networks (GANs) and how do they work? [Video] Video-to-video synthesis method: A GAN by NVIDIA & MIT CSAIL is now Open source Generative Adversarial Networks: Generate images using Keras GAN [Tutorial]

Korean researchers have developed a GAN that can achieve image translation in challenging cases. Instance-aware GAN or InstaGAN as the authors call it can achieve image translation in various scenarios showing better results than CycleGAN in specific problems. It can swap pants with skirts and giraffes with sheeps. The system is designed by a student and an assistant professor from Korea Advanced Institute of Science and Technology and an assistant professor from the Pohang University of Science and Technology. They have published their results in a paper titled InstaGAN: Instance-Aware Image-to-Image Translation, last week. Systems that map images aren’t new but the authors of the paper say that they are the first to report image to image translation in multi-instance transfiguration tasks. The methods used in image to image translation before InstaGAN often failed in challenging cases, like multi-instance transfiguration where significant changes are involved. A multi-instance transfiguration task involves multiple individual objects present in an image. The objective here is to swap objects in an image without changing the background scene. InstaGAN uses the instance information such as object segmentation masks improving on the challenging areas for image to image transformation. The method showed in the paper translates an image as well as its instance attributes. They introduce a context-preserving loss, which encourages the network to learn the identity function outside the target instances. A sequential mini-batch training technique handles multiple instances when using a limited GPU memory. This also enhances the network to generalize better when multiple instances are involved. The researchers compared InstaGAN with CycleGAN and doubled the number of parameters for CycleGAN. This is done for a fair comparison as InstaGAN uses two networks for image and masks. In areas where CycleGAN fails, the new method generates ‘reasonable shapes’. InstaGAN preserves the background while making changes to the objects in images where CycleGAN is unable to maintain the original background. Source: InstaGAN: Instance-Aware Image-to-Image Translation The authors said that their ideas of using the set-structured side information have potential applications in other cross-domain generation tasks such as neural machine translation or video generation. For more details, examples of the model where images are swapped, check out the research paper. NVIDIA demos a style-based generative adversarial network that can generate extremely realistic images; has ML community enthralled Generative Adversarial Networks: Generate images using Keras GAN [Tutorial] What are generative adversarial networks (GANs) and how do they work? [Video]

A pair of researchers, Ruohan Gao, University of Texas and Kristen Grauman, Facebook AI research came out with a method, earlier this month, that can teach an AI system the conversion of ordinary mono sounds into binaural sounds. The researchers have termed this concept as “2.5D visual sound” and it uses a video to generate synthetic 3D sounds. Background According to the researchers, binaural audio provides a listener with the 3D sound sensation that allows a rich experience of the scene. However, these recordings are not easily available and require expertise and equipment to obtain.  Researchers state that humans generally determine the direction of a sound with the help of visual cues. So, they have used a similar technique, where a machine learning system is provided with a video involving a scene and mono sound recording. Using this video, the ML system then figures out the direction of the sounds and further distorts the “interaural time and level differences” to generate the effect of a 3D sound for the listener. Researchers mention that they have devised a deep convolutional neural network which is capable of learning how to decode the monaural (single-channel) soundtrack into its binaural counterpart. Visual information about object and scene information is injected within the CNN during the whole process. “We call the resulting output 2.5D visual sound—the visual stream helps “lift” the flat single channel audio into spatialized sound. In addition to sound generation, we show the self-supervised representation learned by our network benefits audio-visual source separation”, say researchers. Training method used For the training process, researchers first created a database of examples of the effect that it wants the machine learning system to learn. Grauman and Gao created a database using binaural recordings of over 2,265 musical clips which they had also converted into videos. The researchers mention in the paper, “Our intent was to capture a variety of sound-making objects in a variety of spatial contexts, by assembling different combinations of instruments and people in the room. We post-process the raw data into 10s clips. In the end, our BINAURAL-MUSIC-ROOM dataset consists of 2,265 short clips of musical performances, totaling 6.3 hours”. The equipment used for this project involved a 3Dio Free Space XLR binaural microphone, a GoPro HERO6 Black camera, and a Tascam DR-60D recorder as an audio pre-amplifier. The GoPro camera was mounted on top of the 3Dio binaural microphone to mimic a person seeing and hearing, respectively. The GoPro camera records videos at 30fps with stereo audio. Researchers then used these recordings from the dataset for training a machine-learning algorithm which could recognize the direction of sound from a video of the scene. Once the machine learning system learns this behavior, it is then capable of watching a video and distorting a monaural recording to simulate where the sound is ought to be coming from. Results The video shows the performance results of the research which is quite good. In the video, the results of 2.5D recordings are compared against monaural recording.                                                     2.5D Visual Sound However, it is not capable of generating a complete 3D sound and there certain situations that the algorithm finds difficult to deal with. Other than that, the ML system cannot consider any sound source that is not visible in the video, and the ones that it has not been trained on. Researchers say that this method works best for music videos and they have plans to extend its applications. “Generating binaural audio for off-the-shelf video could potentially close the gap between transporting audio and visual experiences, and will be useful for new applications in VR/AR. As future work, we plan to explore ways to incorporate object localization and motion, and explicitly model scene sounds”, say the researchers. For more information, check out the official research paper. Italian researchers conduct an experiment to prove that quantum communication is possible on a global scale Stanford researchers introduce DeepSolar, a deep learning framework that mapped every solar panel in the US Researchers unveil a new algorithm that allows analyzing high-dimensional data sets more effectively, at NeurIPS conference

Earlier this year, a group of researchers from Osaka University introduced an AI system based on convolutional neural network (CNN) for automatically identifying radioresistant tumor cells. In a single cancer tumor, there can be tremendous variation in the cancer cells types which can make it difficult for doctors to make accurate assessments of cell types. Further, this process can be time-consuming and can often be hampered by human error. This AI-based system can make it easy for doctors to choose the most effective treatment and also finds applications in preclinical cancer research. Explaining the utility of this system, one of the researchers, Masayasu Toratani said, “The automation and high accuracy with which this system can identify cells should be very useful for determining exactly which cells are present in a tumor or circulating in the body of cancer patients. For example, knowing whether or not radioresistant cells are present is vital when deciding whether radiotherapy would be effective, and the same approach can then be applied after treatment to see whether it has had the desired effect.” For the study, the researchers used phase-contrast images of radioresistant clones for two cell lines: mouse squamous cell carcinoma NR-S1, and human cervical carcinoma ME-180. They gathered 10,000 images of each of the parental NR-S1 and ME-180 controls as well as radioresistant clones. VGG16, a convolutional neural network for object recognition, was then trained on 8,000 images of cells. For testing the model, the researchers used another 2,000 images to check its accuracy. The model was able to give an accuracy of 96%. As per the results, it had learned the features that distinguish mouse cancer cells from human ones, and radioresistant cancer cells from radiosensitive ones. The features extracted by this trained CNN were then plotted using t-distributed stochastic neighbor embedding, and the plot showed that the images of each cell line were well clustered. In the future, the researchers will train the system on different types of cell types to make it a universal system that can automatically identify and distinguish all variants of cancer cells. To know more in detail, check out the study published by Cancer Research. REVOLVER: A machine learning approach to forecast cancer growth Google, Harvard researchers build a deep learning model to forecast earthquake aftershocks location with over 80% accuracy Stanford researchers introduce DeepSolar, a deep learning framework that mapped every solar panel in the US

A top Amazon executive Anne Rung, advised the Trump administration on the launch of a new internet portal which is expected to generate billions of dollars for Amazon, the Guardian reports. The emails seen by the Guardian indicate ways in which the portal will give the technology giant a dominant role in how the US government buys everything from paper clips to office chairs. Emails exchanged between Rung and the GSA Anne communicated with a top official Mary Davie at the General Services Administration (GSA) about the approach the government would take to create the new portal called as the “Amazon amendment”. This communication took place in 2017, before the legislation that created the portal was signed into law, later last year. According to an The Intercept, the amendment, Section 801 of the National Defense Authorization Act (NDAA), would allow for the creation of an online portal that government employees could use to purchase items including office supplies or furniture. Experts even commented on how this deal would help Amazon establish a tight grip on the $53 billion government acquisitions market. The emails offer new insights into how Amazon has used key former government officials it now employs (directly and as consultants ) to gain influence and potentially shape lucrative government contracts. One of the emails showed the setup of a meeting between Anne and Mary. Rung wrote: “IF the legislation is enacted, I have a sense of how GSA will want to approach this (first you have to select providers, then you will want to implement something incrementally/phased approach), but I want to make sure that I’m not way off the mark. It will help me design a discussion/agenda for our meeting next month.” On asking Davie if they should wait until after the legislation is passed to discuss it, Davie responded that the administration was planning on moving ahead regardless of the outcome of the bill on Capitol Hill. The Guardian also reports that it has not yet been determined which companies will build the US government’s new e-commerce portal, but Amazon is expected to take on a dominant role, giving it a good footing in the market for federal procurement of commercial products. This is not the first time that Amazon has involved itself with the U.S. government. Jeff Bezos, the owner of Amazon and the Washington Post, has had a troubled past with President Donald Trump. However, it seems this hostility is reserved for Trump personally and not the US government. Amazon already operates a cloud service for the US intelligence community, including a contract with the CIA, and has said it can protect even the most top secret data on a cloud, walled off from the public internet. Stacy Mitchell, co-director of the Institute for Local Self-Reliance, a group that supports local businesses said in a statement that “Amazon wants to be the interface between all government buyers and all the companies that want to sell to government, and that is an incredibly powerful and lucrative place to be.” Statements from the GSA and Amazon The Guardian says that Amazon declined to comment on questions related to how much of its business is currently connected to the federal government. However, the tech giant did admit that it had been engaging in “continuous conversations with the GSA” and that it commended the agency for “transforming the conversation around online portals”. Amazon said Rung had been compliant with all White House ethics rules. The GSA said in a statement to the Guardian that during 2017 and 2018, it had met with 35 companies in 2017 and 2018 to discuss “existing commercial capabilities and conduct market research” regarding the e-commerce platforms. The statement says: “No company has been given special access. Instead, all companies expressing interest in the Commercial Platforms program have equal access to GSA. We cannot speculate on which companies will be part of the proof of concept until proposals are received, evaluated, and awards are made.” Trouble for Rung? The Federal law mandates a “cooling off” period of one year before a former senior government official works on projects which he/she has worked on, in the government. It is not clear whether Rung’s communication would be considered a violation of this specific ethics law because there are no details on what exactly she worked on before leaving the government. Lisa Gilbert, the vice-president of legislative affairs at Public Citizen, a consumer advocacy group, said that while she did not believe that the engagement between Rung and Davie was a violation of the law, it was “unsavory” to think that former government officials used their inside knowledge of how the “ballgame” works for their corporate advantage. Gilbert said “There is nothing inherently wrong in talking to stakeholders who will be impacted by the legislation. Our overwhelming worry is that corporate stakeholders have special access that other stakeholders--like public interest groups--do not get.” Mitchell, the small business advocate, said the Rung emails “display an inside relationship that other competing companies don’t have” and show how government infrastructure was being designed with input from Amazon, giving it a big advantage. For more insights on this news, visit The Guardian’s complete coverage. NYT says Facebook has been disclosing personal data to Amazon, Microsoft, Apple and other tech giants; Facebook denies claims with obfuscating press release Amazon Rekognition faces more scrutiny from Democrats and German antitrust probe Sally Hubbard on why tech monopolies are bad for everyone: Amazon, Google, and Facebook in focus  

Microsoft plans to introduce two new webcams next year. One feature is designed to extend Windows Hello facial recognition to all the Windows 10 PCs. The other feature will work with the Xbox One, bringing back the Kinect feature that let users automatically sign in by moving in front of the camera. These webcams will be working with multiple accounts and family members. Microsoft is also planning to launch its Surface Hub 2S in 2019, an interactive, digital smart board for the modern workplace that features a USB-C port and upgradeable processor cartridges. PC users have relied on alternatives from Creative, Logitech, and Razer to bring facial recognition to desktop PCs. The planned webcams will be linked to the USB-C webcams that would ship with the Surface Hub 2, whichwill be launched next year. Though the Surface Hub 2X is expected in 2020. In an interview with The Verge in October, Microsoft Surface Chief, Panos Panay suggested that Microsoft could release USB-C webcam soon. “Look at the camera on Surface Hub 2, note it’s a USB-C-based camera, and the idea that we can bring a high fidelity camera to an experience, you can probably guess that’s going to happen,” hinted Panos in October. A camera could possibly be used to extend experience beyond its own Surface devices. The camera for Windows 10, for the first time, will bring facial recognition to all Windows 10 PCs. Currently, Windows Hello facial recognition is restricted to the built-in webcams just like the ones on Microsoft's Surface devices. According to  Windows watcher Paul Thurrott, Microsoft is making the new 4K cameras for Windows 10 PCs and its gaming console Xbox One. The webcam will return a Kinect-like feature to the Xbox One which will allow users to authenticate by putting their face in front of the camera. With the recent Windows 10 update, Microsoft enabled WebAuthn-based authentication, that helps in signing into its sites such as Office 365 with Windows Hello and security keys. The Windows Hello-compatible webcams and FIDO2, a password-less sign in with Windows Hello at the core, will be launched together next year. It would be interesting to see how the new year turns out to be for Microsoft and its users with the major releases. Microsoft urgently releases Out-of-Band patch for an active Internet Explorer remote code execution zero-day vulnerability NYT says Facebook has been disclosing personal data to Amazon, Microsoft, Apple and other tech giants; Facebook denies claims with obfuscating press release Microsoft open sources Trill, a streaming engine that employs algorithms to process “a trillion events per day”

At the ‘Made by Google’ event, Google launched Pixel 3 sharing its newly introduced camera features, and one of them was Top Shot. Last week, it shared further details on how Top Shot works. Top Shot saves and analyzes the image taken before and after the shutter press on the device in real-time using computer vision techniques and then recommends several alternative high-quality HDR+ photos. How Top Shot works? Once the shutter button is pressed, Top Shot captures up to 90 images from 1.5 seconds before and after the shutter press, and simultaneously selects up to two alternative shots to save in high resolution. These alternative shots are then processed by Visual Core as HDR+ images with a very small amount of extra latency and are embedded into the file of the Motion Photo. Source: Google AI Blog Top Shot analyzes captured images based on three attributes: functional qualities like lighting, objective attributes like whether the people in the image are smiling, and subjective qualities like emotional expressions. It does this by using a computer vision model, an optimized version of the MobileNet model, which operates in low latency, on-device mode. In early layers, the model detects low-level visual attributes like it identifies whether the subject is blurry. In subsequent layers, it detects more complex objective attributes like whether the subject's eyes are open and subjective attributes like whether there is an emotional expression of amusement or surprise. The model was trained using a technique named knowledge distillation, which compresses the knowledge in an ensemble of models into a single model, over a large number of diverse face images using quantization during both training and inference. To predict the quality scores for faces, Top Shot uses a layered Generalized Additive Model (GAM) and combines them into a weighted-average “frame faces” score. For the use cases where faces are not the primary subject, three more scores are added to the overall frame quality scores. These scores are subject motion saliency score, global motion blur score, and 3A scores (auto exposure, autofocus, and auto white balance). All these scores were used to train the model predicting an overall quality score, which matches the frame preference of human raters, to maximize end-to-end product quality. To read more in detail, check out the post on the Google AI Blog. Google launches new products, the Pixel 3 and Pixel 3 XL, Pixel Slate, and Google Home Hub Google’s Pixel camera app introduces Night Sight to help click clear pictures with HDR+ Google open sources DeepLab-v3+: A model for Semantic Image Segmentation using TensorFlow

ProPublica, a non-profit newsroom known for its investigative journalism, published an article yesterday, written by Jeremy B Merrill, their news app developer, on their investigative project around Facebook. They collected over 100,000 targeted Facebook Ads to understand and report how political messaging works on Facebook. It was also meant to analyze how the system manipulates the public discourse. Merrill states that they launched their Facebook Political Ad Collector project in fall 2017 which was joined by over 16,000 people. As a part of the project, all the participants were required to install a browser plug-in that would anonymously send the ads they see on browsing Facebook. As the data was getting collected, it was observed that the number of ads collected from Democrats and progressive groups was larger than from Republicans or conservative groups. “We tried a number of things to make our ad collection more diverse: to start, we bought our own Facebook ads asking people across a range of states to install the ad collector. We also teamed up with Mozilla, maker of the Firefox web browser, for a special election-oriented project, in an attempt to reach a broader swath of users”, writes Merrill. However, since the political ad collector was entirely anonymous, not much information could be gathered about the audience. Another issue was that the left-leaning groups made use of Facebook advertising more than the conservative groups. To solve this issue, ProPublica partnered up with a research firm called YouGov. This was to create a panel of users ranging from a wide spectrum of demographic groups and political ideologies who would be okay with a new less-anonymous ad collector plug-in. A unique ID was assigned to these users that were tied back to the data about them such as demographics, political preference, race, and residence state, which was provided by YouGov. This partnership was funded by the Democracy Fund. YouGov was able to link the answers of the users to demographic questions such as age and partisanship to the ads received. The process of collecting data from the users of the original and publicly available ad collector plug-in that did not participate in the YouGov survey was still the same. These users still remained anonymous to ProPublica. On the other hand, ads that were seen by the participants in the YouGov survey, with their demographic data stripped, became a part of ProPublica’s existing ads database. Learnings from the project After receiving a diverse sample of data regarding the Facebook political ads, ProPublica reached the following conclusions: More than 70% of all the political ads were largely targeted by ideology. Most of these ads were presented to “at least twice as many people from one side of the political spectrum than the other”. Moreover, only about 18% of these political ads were seen by an even ratio of liberals and conservatives. One of the major advertiser targeting both the sides of the political spectrum (liberals and conservatives) was AARP ( American Association of Retired Persons). AARP had spent around $700,000 on ads starting from May to the election. Most of these ads encouraged users to vote. Other ads urged people to hold their member of Congress accountable for voting yes on “last year’s bad health care bills.” The AARP has opposed efforts to replace the Affordable Care Act. One of the ads seen by a majority of the people in the YouGov sample was from Tom Steyer’s “Need to Impeach” organization. This ad included a video saying, “We need to impeach Donald Trump before he does more damage,” for migrant children and Hurricane Maria deaths. The ad was seen mostly by the “self-identified liberals” in the YouGov’s sample. A mysterious Facebook page called “America Progress Now” was discovered by ProPublica that urged liberals to vote for Green Party candidates. “The candidates themselves had never heard of the group, and we couldn’t find any address or legal registration for it”, writes Merrill. A lot of other ads from liberal groups were seen that used misleading tactics similar to the ones used by groups such as Internet Research Agency in Russia to interfere with the 2016 US presidential elections. One such project called “Voter Awareness project” asked conservatives not to vote to re-elect Ted Cruz, the Texas Republican senator, giving examples of Trump’s previous antagonism towards him. This group was however liberal. There were other liberals too such as the Ohio gubernatorial candidate Rich Cordray and the Environmental Defense Action Fund that ran political ads from pages with names of news organizations such as the “Ohio Newswire” and “Breaking News Texas.” Merrill states that although they found out a way to determine how an ad is targeted, there are other complexities to Facebook’s systems which it can’t detect or understand. ProPublica is still looking out for answers to questions such as the impact of the algorithm used by Facebook to show ads to people who are most likely to click, the effect of some people seeing more expensive ads than others, how are cheap ads different from more expensive ones, and so on. ProPublica is currently working on the Ad Collector project and will make future announcements regarding their further studies. For more information, read the official ProPublica Post. Facebook introduces a fully convolutional speech recognition approach and open sources wav2letter++ and flashlight Facebook halted its project ‘Common Ground’ after Joel Kaplan, VP, public policy, raised concerns over potential bias allegations NYT says Facebook has been disclosing personal data to Amazon, Microsoft, Apple and other tech giants; Facebook denies claims with obfuscating press release

Last week, Facebook AI Research (FAIR) speech team introduced the first fully convolutional speech recognition approach. Additionally, they have also open-sourced flashlight, a C++ library for machine learning and wav2letter++, a fast and simple system for developing end-to-end speech recognizers. Fully convolutional speech recognition approach The current state-of-the-art-speech recognition systems are built on RNNs for acoustic or language modeling. Facebook’s newly-introduced system provides an alternative approach based solely on convolutional neural networks. This system eliminates the feature extraction step altogether as it is trained end-to-end to predict characters from the raw waveform. It uses an external convolutional language model to decode words. The following diagram depicts the architecture of this CNN-based speech recognition system: Source: Facebook Learnable frontend: This section of the system first contains a convolution of width 2 that emulates the pre-emphasis step followed by a complex convolution of width 25 ms. After calculating the squared absolute value, the low-pass filter and stride perform the decimation. The frontend finally applies a log-compression and a per-channel mean-variance normalization. Acoustic model: It is a CNN with gated linear units (GLU), which is fed with the output of the learnable frontend. These acoustic models are trained to predict letters directly with the Auto Segmentation Criterion. Language model: The convolutional language model (LM) contains 14 convolutional residual blocks and uses GLUs as the activation function. It is used to score candidate transcriptions in addition to the acoustic model in the beam search decoder. Beam-search decoder: The beam-search decoder is used to generate word sequences given the output from our acoustic model. Apart from this CNN-based approach, Facebook released the wav2letter++ and flashlight frameworks to complement this approach and enable reproducibility. flashlight is a C++ standalone library for machine learning. It uses the ArrayFire tensor library and features just-in-time compilation with modern C++. It targets both CPU and GPU backends to provide maximum efficiency and scale. The wav2letter++ toolkit is built on top of flashlight and written entirely in C++. It also uses ArrayFire as its primary library for tensor operations. ArrayFire is a highly optimized tensor library that can execute on multiple backends including a CUDA GPU and CPU backed. It supports multiple audio file formats such as wav and flac. And, also supports several feature types including the raw audio, a linearly scaled power spectrum, log-Mels (MFSC) and MFCCs. To read more in detail, check out Facebook’s official announcement. Facebook halted its project ‘Common Ground’ after Joel Kaplan, VP, public policy, raised concerns over potential bias allegations Facebook releases DeepFocus, an AI-powered rendering system to make virtual reality more real The district of Columbia files a lawsuit against Facebook for the Cambridge Analytica scandal

Last week, in a study published on JAMA Network Open, researchers revealed that machine learning algorithms trained with physical activity data collected from health tracking devices can be used to re-identify actual people. This study indicates that the current practices for anonymizing health information are not sufficient enough. Personal health and fitness data collected and stored by fitness wearable devices can be potentially sold to third parties, like employers, insurance providers, and other companies, without the users’ knowledge or consent. Also, health app makers might be able to link users name to their medical record and then sell this information to third-parties. Location information from activity trackers could be used to reveal sensitive military sites. Therefore, there is a need for a deidentification algorithm that aggregates the physical activity data of multiple individuals to ensure privacy for single individuals. For this study, the researchers analyzed the National Health and Nutrition Examination Survey (NHANES) 2003-2004 and 2005-2006 datasets. These datasets included recordings from physical activity monitors, during both a training run and an actual study mode, for 4,720 adults and 2,427 children. How does the reidentification procedure work? The machine learning model was constructed by building a separate multiclass classifier for each combination of demographic attributes. They used two different machine learning algorithms for multiclass classification, namely, linear support vector machine and random forests. The models were then tested by feeding in the demographic and physical activity data, but not the record numbers, from the testing data into the models to make predictions of record numbers. The accuracy of the models was calculated by counting how many predicted record numbers matched the actual record numbers in the testing data. The following block diagram depicts the steps of this procedure: Source: JAMA Network Open Results of this study The random forest algorithm was able to reidentify the demographic and physical activity data of 4478 adults (94.9%) and 2120 children (87.4%) in NHANES 2003-2004 and 4470 adults (93.8%) and 2172 children (85.5%) in NHANES 2005-2006. The linear SVM algorithm was able to reidentify the demographic and physical activity data of 4043 adults (85.6%) and 1695 children (69.8%) in NHANES 2003-2004 and 4041 adults (84.8%) and 1705 children (67.2%) in NHANES 2005-2006. How privacy risks can be reduced? Per the research paper, the privacy risks posed on individuals by sharing physical data can be reduced by sharing data not only in time but also across individuals of largely different demographics. This is particularly important for governmental organizations such as NHANES that publicly release large national health datasets. Also, currently we do not have strict regulations for organizations that collect and share these sensitive health data. Policymakers should develop regulations to minimize the sharing of activity by device manufacturers. You can go through the research paper for more details: Feasibility of Reidentifying Individuals in Large National Physical Activity Data Sets From Which Protected Health Information Has Been Removed With Use of Machine Learning. Researchers unveil a new algorithm that allows analyzing high-dimensional data sets more effectively, at NeurIPS conference Researchers develop new brain-computer interface that lets paralyzed patients use tablets Facebook AI researchers investigate how AI agents can develop their own conceptual shared language

Wall street journal published a report yesterday that states Facebook had put a halt on its project named “common ground”, late summer, this year, over the concerns that the project could lead to the accusations of political bias on the platform. Common Ground was developed with an aim to promote healthier political discussions for users with differing political beliefs.   It is reported that the Common Ground Project would have consisted of many different features aimed at reducing the toxic content on the platform and encouraging more positive content surrounding politics. These features included promoting news stories, status updates, and articles shared by people supporting opposite political beliefs. It would also remove any comments and discussions that would promote negativity or hate speech regarding politics. Facebook has already been taking measures to eradicate hate speech and misinformation on its platform as it published a “blueprint” last month, that talks about updating its news feed algorithm. Joel Kaplan, VP, Global public policy, Facebook, raised concerns Facebook had done its research and discussions regarding the project for well over a year before deciding to cancel it. The common ground project was terminated when Joel Kaplan, VP of global public policy at Facebook, raised issues regarding the project. Facebook, however, hasn’t commented anything about the ‘Common ground” project and Kaplan’s reported role in the decision to halt it. Kaplan’s complaint with the project was that, first, the name “Common Ground”, in itself sounds “patronizing”, and second, the project might lead to Facebook receiving criticism from conservative users. A spokeswoman for Facebook told WSJ that Facebook considers it absolutely "essential" to understand the diverse point of views when it comes to creating projects that are meant to "serve everyone”. Kaplan also believed that this attempt to remove polarization, might, in turn, affect the user engagement on Facebook. However, Kaplan was not the only one and Mark Zuckerberg, CEO, Facebook also echoed Kaplan’s beliefs. WSJ also states that Kaplan’s voice has become stronger since the US 2016 presidential elections with him having a say when it comes to making product related decisions at Facebook. The report states that although Kaplan is promoting anti-bias beliefs, he himself has been a part of recent controversies. For instance, Kaplan attended and sat in on the Congressional hearings for Brett Kavanaugh, a then-supreme court nominee, who was accused of sexual misconduct from multiple women. Kaplan's attendance at the hearing led to a wide outrage from the Facebook employees.  Another example presented in the report is Kaplan’s partnership with the “Daily Caller’s fact-checking entity” that ended in November when “the Daily Caller’s fact-checking operation lost its accreditation”, reports the WSJ. Nothing can be commented on whether Facebook’s decision to halt the project was a wise one or not, however, the fact that Facebook is taking initiatives towards promoting healthier conversations on its platform seems certainly credible. The story first appeared on WallStreet Journal. NYT says Facebook has been disclosing personal data to Amazon, Microsoft, Apple and other tech giants; Facebook denies claims with obfuscating press release Ex-Facebook manager says Facebook has a “black people problem” and suggests ways to improve UK parliament seizes Facebook internal documents cache after Zuckerberg’s continuous refusal to answer question

22nd December marked a win for U.S. government in terms of efficiency, accountability, and transparency of open data. Following the Senate vote held on 19th December, Congress passed the Foundations for Evidence-Based Policymaking (FEBP) Act (H.R. 4174, S. 2046). Title II of this package is the Open, Public, Electronic and Necessary (OPEN) Government Data Act, which requires all non-sensitive government data to be made available in open and machine-readable formats by default. The federal government possesses a huge amount of public data which should ideally be used to improve government services and promote private sector innovation. According to Data Coalition, "the open data proposal will mandate that federal agencies publish their information online, using machine-readable data formats". What does the bill mandate? There are a number of practical things the bill will do, which should have real benefits for both citizens and federal organizations: Makes Federal data more accessible to the public, and requires all agencies to publish an inventory of all their "data assets" Encourages government organizations to use data to make decisions Ensuring better data governance by requiring Chief Data Officers in Federal agencies After some minor corrections made on Saturday, December 22nd, the Senate passed the resolution required to send the bill onwards to the president’s desk. There are two things which were amended in this act before passing it on to the president: The text was amended so that it only applied to CFO Act agencies, not the Federal Reserve or smaller agencies. There was acarve-out “for data that does not concern monetary policy,” which relates to the Federal Reserve, among others. Why is the open data proposal required? For many years, businesses, journalists, academics, civil society groups, and even other government agencies have relied on data that the federal government makes freely available in open formats online. However, while many federal government agencies publish open data, there has never been a law mandating the federal government to do so. The data available in a machine-readable format and catalogued online will help individuals, organizations, and other government offices to use it while preserving privacy and national security concerns. Open data has been an effective platform for innovation in the public sectors supporting significant economic value while increasing transparency, efficiency, and accountability in government operations. It has worked towards powering new tools and services to address some of the country’s most pressing economic and social challenges. Michele Jolin, CEO and co-founder of Results for America, said in a statement. “We commend Speaker Ryan, Senator Murray and their bipartisan colleagues in both chambers for advancing legislation that will help build evidence about the federally-funded practices, policies and programs that deliver the best outcomes. By ensuring that each federal agency has an evaluation officer, an evaluation policy and evidence-building plans, we can maximize the impact of public investments.” U.S Citizens also called this bill a big ‘milestone’ in the history of the country and accepted the news with vigor. https://twitter.com/internetrebecca/status/1076226160751726592 https://twitter.com/Jay_Nath/status/1076884756426457088 You can read the entire backstory on what’s in the bill and how it was passed at E Pluribus Unum. Equifax data breach could have been “entirely preventable”, says House oversight and government reform committee staff report Consumer protection organizations submit a new data protection framework to the Senate Commerce Committee Furthering the Net Neutrality debate, GOP proposes the 21st Century Internet Act

John Giannandrea joined Apple in April 2018, after an 8-year long stint in Google. Yesterday, Apple announced that he has been promoted as the senior vice president of Machine Learning and Artificial Intelligence strategy and moved to the company’s executive team. He will directly report to Apple CEO Tim Cook. “John hit the ground running at Apple and we are thrilled to have him as part of our executive team,” said Tim Cook, Apple’s CEO. “Machine learning and AI are important to Apple’s future as they are fundamentally changing the way people interact with technology, and already helping our customers live better lives. We’re fortunate to have John, a leader in the AI industry, driving our efforts in this critical area”, he added. John will be continuing to look after Apple’s virtual assistant Siri and the Core ML, and Create ML software that developers can use to incorporate artificial intelligence capabilities into their applications. At Google, John was overseeing Google search, along with machine intelligence and research. Apple is facing serious competition from its rivals about incorporating artificial intelligence in their software. Siri, their virtual assistant has been criticized for its shortcomings in comparison to AI offerings from companies like Microsoft, Amazon, and Google. Perhaps this is why Apple has made this move to reorganize its AI teams into a single business unit under John’s leadership and also give him full charge of Siri. https://twitter.com/fromedome/status/1075982956966088704 https://twitter.com/stevekovach/status/1075816852528480256 For additional information, you may visit Apple Newsroom. Apple’s security expert joins the American Civil Liberties Union (ACLU) Apple app store antitrust case to be heard by U.S. Supreme Court today Apple has quietly acquired privacy-minded AI startup Silk Labs, reports Information

Slack has become a mainstay of many industries - when it goes down, you can be sure you'll know about it. However, for a number of users, Slack access appears to have been revoked. Most of these users have an Iranian background. Mahdi Saleh, a PhD student at the Technical University of Munich, explained on Twitter how Slack had terminated his account "in order to comply with export control and economic sanctions laws and regulations promulgated by the U.S. Department of Commerce and the U.S. Department of Treasury." https://twitter.com/mahdimax2010/status/1075524107847000064 Slack responded quickly on Twitter, explaining that "our systems may have detected an account on our platform with an IP address originating from a designated embargoed country." The company then offered to investigate the issue in detail for Saleh. Saleh said that following the exchange he got in touch with Slack, but has not, at the time of writing, heard back from the company. Other Slack users have had their accounts terminated Saleh says that he does not believe his case is unique: "Apparently I am not the only person outside Iran that this happened to" he said. "A lot researchers [sic] and emigrants got the same email from Slack." A quick Twitter search indicates that this is the case, with a number of users sharing the same message from Slack as the one received by Saleh. "I’m a PhD student in Canada with no teammates from Iran!" said Twitter user @a_h_a. "Is Slack shutting down accounts of those ethnically associated with Iran?! And what’s their source of info on my ethnicity?" https://twitter.com/a_h_a/status/1075510422617219077 The same user also called into question whether the reasons given by Slack are true or accurate. "I’m in Canada. No ties to Iran. No teammates in Iran!" "I wonder how my account was associated with my ethnicity and how/where they digged [sic] this info from," he said. Amir Odidi, a software developer at ipinfo.io said that there was "no way to appeal this decision. No way to prove that I'm not living in Iran and not working with Iranians on slack. Nope. Just hello we're banning your account." https://twitter.com/aaomidi/status/1075621119028314112 These aren't the only cases - there are a huge range of other examples of Iranians based in the U.S. and Canada having their accounts terminated. https://twitter.com/nv_rahimi/status/1075695124561125377 https://twitter.com/bamaro_ir/status/1075667601878061056 "Filter coffee, not people", said one user. Slack responds It's hard to say exactly what's going on. We approached Slack to get their perspective; the company provided us with a statement very similar to the one sent to those users who have had their accounts terminated. It reads: “Slack complies with the U.S. regulations related to embargoed countries and regions. As such, we prohibit unauthorized Slack use in Cuba, Iran, North Korea, Syria and the Crimea region of Ukraine. For more information, please see the US Department of Commerce Sanctioned Destinations, The U.S. Department of Treasury website, and the Bureau of Industry and Security website. “Our systems may have detected an account and/or a workspace owner on our platform with an IP address originating from a designated embargoed country. If our systems indicate a workspace primary owner has an IP address originating from a designated embargoed country, the entire workspace will be deactivated. If someone thinks any actions we took were done in error, we will review further.” What does this tell us about how Slack handles user data? With no clear response from Slack, it isn't exactly clear how this happened. You could take Slack at their word, but given the information given by users on Twitter, there does appear to be a piece missing in this puzzle. However, we probably can say that Slack does have an extensive record that allows them to link accounts to specific countries - whether that's via IP address or something else. As one Twitter user wrote, the story suggests that Slack has "more data than some customs and border agencies." https://twitter.com/rakyll/status/1075691304896606208 This article was updated 12.10.2018 10.25am EST to include Slack's response.

It was back in March this year when a self-driving car by Uber killed a pedestrian, a 49-year-old Elaine Herzberg, in Tempe, Arizona. Uber, who had to halt the on-road testing of its autonomous vehicles after the incident, got the permission granted again to restart the testing yesterday. The authorization letter by the Pennsylvania Department of Transportation (PennDOT) confirmed that Uber will resume its on-road testing of self-driving cars in Pittsburgh. As per the details of the accident’s investigation, Rafaela Vasquez, the backup driver, had looked down at his phone 204 times during a course of a 43-minute test drive. After the accident, Uber had to halt all of its autonomous vehicle testing operations in Pittsburgh, Toronto, San Francisco, and Phoenix. Additionally, a shocking revelation was made last week by an Uber manager, Robbie Millie, who said that he tried to warn the company’s top executives about the danger, a few days before the fatal Arizona accident. According to Robbie Miller, a manager in the testing-operations group, he had sent an email to Uber’s top execs, where he warned them about the dangers related to the software powering Uber’s prototype “robo-taxis”. He also said that he warned them about the human backup drivers in the vehicles who hadn’t been properly trained to do their jobs efficiently. Other than that, Uber recently released its Uber safety report, where the company mentioned that it is committed to “anticipating and managing risks” that come with on-road testing of autonomous vehicles, however, it cannot guarantee to eliminate all of the risks involved. “We are deeply regretful for the crash in Tempe, Arizona, this March. In the hours following, we grounded our self-driving fleets in every city they were operating. In the months since, we have undertaken a top-to-bottom review of ATG’s safety approaches, system development, and culture. We have taken a measured, phased approach to return to on-road testing, starting first with manual driving in Pittsburgh”, said Uber. Although Uber has not released any details on when exactly it will be restarting its AV’s road testing, it says that it will only go back to on-road testing when it has implemented the improved processes. Moving on forward, Uber will make sure to always have two employees at the front seat of its self-driving cars at all times. There’s also going to be an automatic braking system enabled to strictly monitor the safety of the employees within these self-driving cars. Uber’s new family of AI algorithms sets records on Pitfall and solves the entire game of Montezuma’s Revenge Uber announces the 2019 Uber AI Residency Uber posted a billion dollar loss this quarter. Can Uber Eats revitalize the Uber growth story