If all you have is a hammer, everything looks like a nail .
Artificial Intelligence (AI) and Machine Learning (ML) once again have become big topics. Massive amounts of investments, overcrowded AI-classes at Stanford, MIT and other leading educational institutions, abundance of AI-related job openings, all these have become a norm these days. A couple weeks ago this article on The Motley Fool even urged investors to keep at least one AI-related stock in their portfolios. The information technology field is already over-complicated and AI makes it even more cryptic. In this article I will try to demystify the field of AI without relying on complex technical details let alone any machinery from mathematics. Accept this as a layman’s intro to Artificial Intelligence. Thus, main audience shall be business people and especially investors considering investing in AI-startups.
A disclaimer is due here. This article will be critical to but objective about what’s really going on in the field. This may surprise you as my startup, AlgorX, Inc. has just started raising seed investment. But in business honesty shall be above all. I hope this article will help you make a better investment decisions and set realistic expectations about AI.
The start of the modern wave of Artificial Intelligence and more specifically Machine Learning can be attributed to the works of Geoffrey Hinton, a professor from the university of Toronto. Hinton’s software practically has beatean everything existing before in image recognition. Believe it or not, computers were suddenly able to recognize images with a 90+ percent accuracy. Not so bad results from a silicon chip indeed.
Main problem what I see here is that this approach is not how our brains work. Just a quick example. If you showed an image with black and white stripes to a computer powered by this software, the image would be confused with that of a zebra. A yellow and black striped image is perceived by the same software as a school bus. Now, take a toddler. She may not say that on one photo she sees a French bulldog and on the other a German shepherd but she will definitely not confuse a dog with a piece of clothes. However, this shall not come as a surprise. Modern AI-software is based on mathematical models that simplify reality. Artificial Neural Networks (ANN) used heavily in modern AI and ML try to model neurons in our brains but only partially. The other part that Artificial Neural Nets don’t capture is considerable. (I promised no cryptic details but for curious minds I will add that, practically all Artificial Neural Nets model only connections between neurons but according to modern neuroscience, apart from connections between neurons there are certain gases that trigger other types of interactions between the neurons. Obviously, this topic is beyond the scope of this article and to be honest, modern neuroscience does not have answers to many questions).
Another problem with modern approach, which is heavily based on a type of Machine Learning called Deep Learning (DL), is that nobody really understands how it does the trick. To be short, modern machine learning and especially deep learning are like blackboxes. On the one hand, this is good because it allows programmers to use machine learning tools without actually knowing what’s going on inside. On the other hand, this is a weakness. Without knowing how a thing works, you will never really achieve the best results. Think about an amateur driver and a professional racing driver. Both drive cars but a casual driver may know nothing about engines, gears, and pistons and most of us don’t. However, a pro driver does knows pretty much about the inner working of the vehicle’s systems and that’s one of the reasons why a pro driver practically always outperforms an amateur. Why is this fact so important for investors? Because investments are made in people not in products and investors shall have a clear picture that if a person is able to use a certain software package for AI or ML it does not necessarily mean that he or she understands what’s happening inside.
Before going any further, let’s clarify the relation of AI to ML and to DL. Put is simply, AI is the whole field. ML is just an approach to AI, one (but not the only) direction. DL is a sub-field, a more narrow specialization of ML. For completeness it must be noted that with very few exceptions all these approaches are quite old. The only reason why modern AI and ML did not take say 20 or 30 years ago is the lack of computing power. Machine Learning and especially Deep Learning requires massive amounts of computing power. Hence, we shall see all kinds of specialized hardware coming to the market. At least Nvidia has clearly declared this and Google TPU is a clear-cut confirmation that we will see more and more specialized hardware. Thus, investing in AI-hardware startups might be quite a smart idea.
Is AI over-hyped? It certainly is. I believe it will not take much time before the current approach faces limits. This may happen in the next five years or the field may grow for a decade. Obviously, by increasing computing power better results can be achieved but a quick example is enough to demonstrate the weakness of this approach. The brain of a typical fly contains 20 thousand or so neurons. Even for the best artificial neural networks to achieve the same result it takes computing power fitting roughly in a size of a modern refrigerator. Definitely you can’t make a refrigerator fly. Again, the nature outsmarts us by a huge margin.
If you look carefully at current trends, AI is trying to solve problems that our brains do on autopilot — vision, voice recognition, speech synthesis, etc. Undoubtedly, these are vital functions but that’s something that our animal friends can do as well (except speech synthesis but some animals like parrots can do that trick). It does not take much intelligence to see things or hear sounds. Millions of years of evolution have taken great care to make these functions practically autonomous. However, when it comes to thinking capabilities this is where the real magic happens. Vision is a miracle in its own right but our ability to understand and process language is an even more fascinating phenomenon. Cognitive psychologists know that the part of the mind that is in charge of language processing, analysis, creativity (and generally what we call thinking) is orders of magnitude slower than the other part that runs on an autopilot. In this direction of AI there is not much progress. Let me suggest a quick experiment. Take your phone and try to talk to your AI-powered assistant (Siri on iPhone and Google Assistant on Android). Do you notice that practically everything ends in Google search? Put it simply, neither Siri, nor Google assistant nor Amazon Alexa are currently able to maintain deep and meaningful conversations with you even though they understand your speech, thanks to modern Deep Learning, with 97% accuracy. Another clear example is machine translation. This problem is very far from being solved. A person with intermediate language knowledge can do much better translation than the state of the art machine translation software allows. Knowledge acquisition, representation, processing and conversation are the directions that we are tackling at AlgorX, Inc. and this is what AI researchers call AI-complete problem. In other words, creating Artificial Intelligence as complex as we, humans are. This is a truly fascinating area of Artificial Intelligence.
To summarize, the field of modern AI is turbulent but progressing. If you are considering an investment in an AI or ML company I recommend that you clearly understand what you are investing in and take whatever you read with a grain of salt. Many AI and ML companies will have huge exits and can deliver enormous profit margins to their investors. But if the whole field does not deliver on the promise, the second AI winter is warrantied.