Mar 09 2020

Using Neural Networks for Image Sensing

A new study published in Nature details the use of a neural network on a 2-dimensional computer chip that by itself can be trained to recognize specific images within nanoseconds. This is more of a proof of concept than something with direct immediate applications, but let’s talk about that concept.

To back all the way out – evolution represents hundreds of millions of years of tinkering with multi-cellular structures, and even longer when talking about biochemistry. This is a natural laboratory that has developed some elegant designs, and at the very least can serve as a useful source of inspiration for modern technology. That is the concept of neural networks, designing computers to work more like a vertebrate brain. Specifically, the “neurons” in a neural network are not just binary, on or off, but rather can fire with various degrees of strength. Further, their firing affects the activity of those neurons they are connected to. Computer hardware with networks designed on these basic principles are called  artificial neural networks (ANN). They hold the promise of not only faster and more powerful computing, but are designed to learn (which is why they are so often associated with artificial intelligence).

Another principle at work here is top-down vs bottom-up processing, another concept that has increasingly been incorporated into AI. If we go all the way back to the early days of AI the basic idea was to create high level computer intelligence that could solve problems with the top down, with deep understanding. That goal, now referred to as general AI, is still a ways off. But meanwhile AI has advanced considerably through more of a bottom-up approach, using algorithms to sift data in increasingly sophisticated and adaptable ways. We now have deep learning AI and other specific processes that can produce impressive results without any general AI “understanding” what it is doing.

One question is – will we be able to build a general AI out of these limited AI components? Is it just a matter of building in enough sophistication and complexity? We won’t know until we do it, but if living organisms are any guide, I think there is reason to be positive. Specifically – that is basically how our brains work.

Most of the cognitive processes that our brains do are more like limited AI than general AI – they are task specific and subconscious. Taking any task, like walking. Your top-level, conscious, cortical self is not, top-down, controlling all the movements necessary to walk. Walking is very much a limited AI type of task. Your vestibular system is sensing the direction of gravity and acceleration and feeding that information to your cerebellum, which also takes in visual information and feedback from your muscles to produce coordinated movement. You further have dedicated “anti-gravity” systems in your brainstem that automatically produce muscle tone so as to maintain posture upright against gravity. When you walk all you really have to think about consciously is where you want to go, and maybe some navigational strategy. You don’t have to worry about the basics.

As an analogy – image two video games. In one game you control an avatar completely. You can operate a number of muscles by hitting keys and moving your mouse, and with those controls you have to move each leg, maintain trunk position, and use the visual feedback of what your avatar is doing in order to maintain control. Imaging walking with that system – it would be horrifically complicated, and probably beyond the control you could exercise with a keyboard.

In the second video game, which operates like most actual video games in existence, you hit the WASD keys to walk in the desired direction, but the walking is all automatic. This latter video game is much more like your actual brain than the former game.

The general biological intelligence of our brains is built upon many limited biological intelligence algorithms that are task specific. You don’t think about breathing, or the muscle coordination necessary for swallowing. You can keep your eyes focused on a target while moving your head. Even something sophisticated like speaking – you don’t think about the specific tongue and mouth movements you make, only about the thoughts you want to express. (Of course, when these subsystems break down, people may think about them a great deal.)

So – when we want to design a robot that can walk, we don’t need to design a general AI robot that can think. We just need to duplicate, essentially, the specific AI subsystems that subconsciously produce things like anti-gravity balance and locomotion. We can build behaviors from the bottom up with narrow, but efficient, dedicated algorithms.

Now let’s turn back to the current study. They are using an ANN that is connected to a computer which alters the function of the ANN so that it will learn to recognize specific images. However, once it is trained on those images, the computer can be disconnected, and the ANN can operate by itself to distinguish a limited set of images in nanoseconds. This is a very limited function, but just like the processing components that build the general intelligence of vertebrate brains.

The immediate applications of this kind of visual processing are mostly industrial and research, and even then it will have to be scaled up from what was used in this proof of concept. The authors specifically mention recognizing fracture formation in structures as an example, or recognizing specific letters. The power in this approach is the speed, which is near-instantaneous. And that, of course, is exactly why biological networks evolved the way they did. We can make fast microadjustments to our muscle tone to maintain balance against gravity because our complex thinking does not need to be involved. Efficient brainstem networks do it automatically.

Beyond the specific applications mentioned, I think if this approach works well, it or something like it can be incorporated into more sophisticated AI visual processing systems. Again – this is how our visual system works. Vertebrate vision is processed at every level in the circuit, from the retina to the highest cortical levels. At each step learning algorithms are used to do all sorts of basic processing, such as interpreting color, emphasizing contrast, and making adjustments for shading. At higher levels too, our visual system makes assumptions about size and distance, matches patterns to known things and then enhances the processed image to make a better match. Our visual system even makes a time correction to compensate for the delay all this processing takes.

Most of this visual processing takes place subconsciously with fast and efficient dedicated neural networks. I can easily imagine the ANN in this study being incorporated into a more general visual processing system, doing some fundamental preprocessing in nanoseconds to speed up the whole process. This study is one tiny step in this direction, but this is one plausible direction that AI can go in.

The ultimate question is – if you keep doing this, using ANNs built for specific functions, feeding into increasing more powerful and sophisticated AI systems – at some point will you have something that basically operates like a general AI? I don’t know, but the fact that this is exactly how evolution produced humans may be some indication.

No responses yet