Artificial Intelligence: From Neural Systems to Neural Networks, by James Wadman
This is a very casual meditation on the state and future of artificial intelligence from the perspective of a neurobiologist.
I want to clarify that I will use two different phrases to make it easy to distinguish between biology and computer science. In this piece, “neural systems” refers to the biological nervous system and the associated molecular biology. “Neural networks” refers to artificial intelligence that mimics (for now) neural systems on the macro-scale of neurons. I will argue that neural networks should operate on a more fundamental scale, where instead of focusing on macro-scale neurons, we should focus on the underlying molecular-scale (genetic) processes. To address the hardcore physicists: yes, we can and will at some point go one step further and operate neural networks on an atomic (dare I say quantum) scale, but this will come only after we fully understand molecular-scale neural networks. More on this later.
First let me ask, what are you trying to accomplish with designing an artificial intelligence system? One of the more common applications of artificial intelligence is to compose an algorithm that can evaluate and improve itself. Let’s call this AIa. This utilizes simplified (although not necessary simple) versions of neural networks that take into consideration the connectivity between neurons using a circuit-structured and feedback mentality. Another form of artificial intelligence includes algorithms that can make sense of large, often abstract, data structures and draw conclusions from the chaos that humans would not see. Let’s call this AIb. This form might also use neural network feedback tools, but focuses on data rather than the plasticity of the algorithm. Both of these take simple human logic characteristics, such as deduction and compare/contrast, and apply them to data that is far too vast for an ordinary human to sift through. Either of these examples are only as complex as the data itself or the results desired. In other words, the more abstract the data or the more precise the conclusions, the more challenging the algorithms. The differences between the two can be understood by comparing them to behavioral traits in humans. While AIa is based on the principles of learning (the brain adapts to experience to create intelligence), AIb is mimics intuition (the brain extrapolates direct conclusions from abstract concepts). However, one can write these algorithms without a fundamental understanding of how actual biological intelligence works.
If, however, our goal for artificial intelligence is to make a true intelligent entity we have a major problem. We should not expect to fully replicate human intelligence or consciousness through neural networks until we fully understand neural systems. Ironically, we also cannot fully understand neural systems without our work so far in using artificial intelligence algorithms in the manner of the first two examples I discussed above. AIa and AIb are integral tools to understanding the complexity of neural systems with incredible capacity and utility in modern computing. But they are not yet acceptable for making true consciousness.
In the wake of my seemingly pessimistic sentiments, I want to make clear that our only barrier to artificial intelligence is our knowledge of the problem itself and the complexity of our algorithms. Once we really understand what we are doing, the solution will be obvious. For now, we should be working toward optimizing functional use algorithms (AIa and AIb) to build the foundation for true intelligence systems.
AIa and AIb are powerful algorithms for modeling molecular biology and are therefore the keys to understanding the problem, not the solution. Neural networks are designed to solve problems or enhance algorithms by simulating neuron-neuron interaction and plasticity. Let us not power to quickly forward in attempts to create functional artificial intelligence algorithms before truly understanding the capacity of learning and ingenuity. Concisely stated, the solution to creating an algorithm for artificial consciousness is to create realistic neural networks based on the underlying molecular neural systems. Go beyond just neuron plasticity, incorporate genetics and cell signaling, and build upon neuron replication.
Artificial intelligence has so far undervalued multi-stage memory consolidation and altogether neglected the ironic cohesion between the unconscious and conscious minds. We are therefore taking blind shots into a darkened room that is likely far larger than the echoes that call back. This comes as a contrast to my often optimistic perspective on technology and our ascension toward the future, so what is my intention in saying this? Quite simply I am suggesting that more biologists should program and more programmers should study the brain’s fundamental biology. Everything in biology is a product of the molecular processes that govern the expression, activation, and inhibition of genetics. It would be foolish, therefore, to limit neural networks to neuron-neuron interactions when their actual function involves far more complicated interactions. Human intelligence might be a function of synaptic interactions, but we cannot ignore the fact the synaptic interactions are the function of intracellular molecular processes. At the present moment, we are too limited by our knowledge of what happens “under the hood” of neurons to properly model intelligence through a synaptic model. We should develop accurate libraries for how biological systems work to the point where neural networks in computer science are equally complex and efficient as neural systems in human conscious reflections, unconscious reflexes, and intelligence.
It is already inherent to computer algorithms that the efficiency of problem solving far exceeds the efficiency of a human with a pen and paper. That is to say, given any complex question and a computer program with the algorithm and a person to knows how to solve the problem, the computer program will nearly always win. Artificial intelligence, even in its early stages of simply replicating human thought, will create far more capable “thinkers.” However, artificial intelligence entities will not be restrained by genetics, evolution, and neurochemistry. We will be able to take strides from there to improve upon the biological disadvantages of natural neural system to the point where we will eventually determine strategies for enhancing the basic framework of thought beyond that of the capacity of humans.
I believe that the future of artificial intelligence will progress in stages that address both our strides so far and the distance left before us to the end goal. Its success will be based on genuine understanding and the pursuit of knowledge not driven by profit or social gain. However, in the end artificial intelligence will be profitable, but the benefits beyond just currency are what will make the true difference. Our moral, intellectual, and sensible views of modern structure will be forever improved.