Chapter 11 – Conceptual Representation Learning

1. The curse of dimensionality leads to reducing dimensions and features in machine learning algorithms. (Yes | No)

Yes. The volume of data and features makes it necessary to extract the main features of an observed event (an image, sound, and words) to make sense of it.

Overfitting and underfitting apply to dimensionality reduction as well. Reducing the features until the system works in a lab (overfitting) might lead to nowhere once the application faces real-life data. Trying to use all the features might lead to underfitting because the application solves no problem at all. 

Regularization applies not just to data but to every aspect of a project.

2. Transfer learning determines the profitability of a project. (Yes | No)

Yes if an application of an AI model in itself was unprofitable the first time but could generate profit if used for a similar type of learning. Reusing some functions would generate profit, no doubt.

No, if the first application was extremely profitable but "overfitted" to meet the specifications of a given project.

3. Reading model.h5 does not provide much information. (Yes | No)

No. Saving the weights of a TensorFlow model is vital during the training process to control the values. Furthermore, trained models often use HDF files (.H5) to load the trained weights.  A Hierarchical Data Format (HDF) contains multidimensional arrays of scientific data.

4. Numbers without meaning are enough to replace humans. (Yes | No)

Yes. In many cases, mathematics provides enough tools to replace humans for many tasks (games, optimization algorithms, and image recognition).

No. Sometimes mathematics cannot solve problems that require concepts such as many aspects of NLP.

5. Chatbots prove that body language doesn't mean that much. (Yes | No)

Yes. In many applications, body language does not provide additional information. If only a yes or no answer is required, body language will not add much to the conversation.

No. If emotional intelligence is required to understand the tone of the user of a chatbot, a webcam detecting body language could provide useful information.

6. Present-day ANNs provide enough theory to solve all AI requests. (Yes | No)

No. Artificial Neural Networks (ANN) cannot solve thousands of problems, for example, translating poetry novels or recognizing images with forms that constantly vary.

7. Chatbots can now replace humans in all situations. (Yes | No)

No. Concepts need to be added. The market provides all the necessary tools. It will take some years to be able to speak effectively with chatbots.

8. Self-driving cars have been approved and do not need conceptual training. (Yes | No)

Yes, that could be true. Sensor, mathematics (linear algebra, probabilities) might succeed within a few years.

No. Certain problems will require concepts (and more robotics) when emergency situations that require creative solutions occur. If a self-driving car encounters a wounded person lying on the road, what is the best approach? The choices are to call for help, find another person if the help arrives too late, pick up the victim, drive them to a hospital (robotics), and much more.

9. Industries can implement AI algorithms for all of their needs. (Yes | No)

Yes. All the tools are there to be used. If the right team decides to solve a problem with AI and robotics, it can be done.

No. Some tools are missing, such as real-time management decision tools when faced with unplanned events. If a system breaks down, humans can still adapt faster to find alternative solutions to continue production.