|What is Intelligence? Pathways to Synthetic Intelligence|
If you follow current AI Research then it will be apparent to you that AI research, the deep learning type has stalled! This does not mean that new areas of application for existing techniques are not appearing but that the fundamentals have been solved and things have become pretty standardized.
So what is the way forward? Deeplearning and other machine learning methods will continually be applied in many areas and in new ways. Many researchers will come out with new tricks and hacks to squeeze out more performance from existing systems but the state of the art of deeplearning systems will not continue to bring out fantastic results like they once did and they will not lead us to the dream of strong AI (The kind of AI you see in movies).
We will see a lot of Human Centered AI applications in various domains but that is all we can hope for. All the advancement current AI techniques have brought us are good and well but they are far from that kind of intelligence that we humans can call a Strong AI. Voice assistants like those from Google, Amazon and Microsoft are all very cool products and have a lot of masterly software engineering behind them. But the thing is, they are based on stretching the capabilities of the current paradigm to the maximum and in order to make progress we need a new paradigm where we can start doing very complex stuff with as little plumbing and "engineering" as usual.
The AI community possesses some of the smartest people on earth and of course, they are not resting on the laurels of the achievements of deep learning. They are actively seeking new means and methods to help us go above the plateau of deep learning that we are currently stuck in.
This can be observed from the recent interest in the works of Alan Turing and most especially is B-Type neural networks published in 1948. This "openness" to new techniques indicates the trend that AI researchers are beginning to look elsewhere for the next thing that will take us out of the well standardized current paradigm of deeplearning.
To go on talking about what the next step towards Strong Artificial Intelligence should be would take an entire book and that is why I decided to do that with my new book: What is Intelligence? Pathways to Synthetic Intelligence. In this book, I open the mind of the researcher to the most promising directions AI research can take towards creating a future where we have a Strong Artificial Intelligence that will take up most of our intellectual burden the same way deep learning, modern computing and even the humble calculator has done in the past.
Yes! Networks are the way forward forever because networks, anything that can be modelled as nodes and edges are fundamental to the very nature of the universe. I push forward the idea that the success of deep learning, a techniques based on exploiting the structure of a network by modifying weighted connections, is only the first step on the pathway to discovery and that we will avail ourselves of much good if we research generalized network structures of all kinds until we discover systems that could replace our current paradigms.
Stephen Wolfram in his book A New Kind of Science explores the idea of Space as a Network which is very interesting because if one looks closely everything that sends information from one point to another can be modelled as a network.
The main purpose of my book: What is Intelligence? Pathways to Synthetic Intelligence is to announce to the research community and everyone interested that we should look at Networks in a more explicit manner than we have done. Thanks to the name "Neural Networks", at least we always have it at the back of our mind when we are doing deep learning that we are really dealing with a network.
Without the title of Neural Networks, we would be lost in all the technicalities of matrix multiplication which is the representation that deep learning systems take on when running on some practical computer. If you were to be handed a deep learning system, it would be hard for you to know that you are dealing with a system that is modelled as a network with weighted connections, just like you see in the neural network diagrams. You will be so focused on the concrete implementation as matrix multiplication such that the abstract representation of the system as a network will elude you.
If you look from above you could clearly see that almost any system can be modelled as a network: from systems of equations to blood flow in the circulatory system to hormone signalling, etc. But when you are in the mud of dealing with the practicalities of a system you would usually use models that are as close to the representation of the problem as possible rather than think in its most abstract representation as a network.
My core argument in the book is that if you are able to find a suitable network representation for some system, then solving problems in that domain will be a matter of manipulating the structure of that network from a problem state to a solution state. This is a very abstract description of what is actually going on with our intelligence and that is why the brain is a very explicit network. I try my best to bring to light most of these ideas and the observant reader will be able to see a clear and deep picture of the whole scheme by the end of the book.
We are on the right path with all the progress in deep learning, but like many have said, we have been trudging blindly towards our destination. With this book, I hope to clear our eyes and show us what exactly we have been doing and with this clarity comes more speed.
There is always the looming issue of AI ethics and while I am not an expert on this, there are many people working towards assuring humanity of a healthy outcome. Many advocate that we should slow AI progress but I do not agree with this view. AI progress is already slow because current techniques consume a lot of resources to produce their results. Slowing down AI research because we want to be more prepared for the arrival of Strong AI is akin to slowing down electric car development because we want to be prepared for the economic consequences of a world where the demand for oil is less, ignoring the fact that internal combustion is tearing our planet apart.
Rather AI ethics people should accelerate their own activities because if the ideas in a book like mine and other similar ideas should be adopted the progress of AI development will start another exponential race to the top.
Current AI ethics work is mostly centred on malicious use of AI technologies by humans on other humans but the future of AI ethics will be more like trying to control nuclear proliferation. Malicious? yes, but far more potent than today's deepfakes.
If one observes the outflow of AI papers it has become mostly one hackery after another, it is clear that the forward direction is not very clear anymore like it was in the very early days where deeplearning started besting humans in image recognition. But we should not despair, the next wave is just around the corner if we are able to look ahead rather than get stuck meddling with techniques that are only examples of much greater things to come.