Make More Efficient AI By Learning from Our Brain Again

As we all know, the recent release of ChatGPT-4 took the world by storm. Its predecessor has already been impressive, but this new version brings our experience with AI (Artificial Intelligence) to a new level. We have come a long way to get machines to think and communicate more like human beings. As a matter of fact, the AI systems represented by ChatGPT are superior to us in many aspects. However, the success of AI also causes people to have mixed feelings about the technology. On the one hand, it is exciting to have a powerful assistant on our side. But on the other hand, it is natural to feel a little humbled and perhaps threatened. How can we control something that is more powerful than us forever? More than 1000 dignitaries including Elon Musk and Steve Wozniak signed an open letter asking for an immediate pause of AI development. The concern is definitely warranted, and it is imperative to get the government and relevant experts together to come up with some solution fast.

The successful AI systems are actually inspired by the human brain. Two research scientists from the University of Chicago, who later moved to MIT, came up with the idea of an artificial neural network in the mid-1940’s. The network, modeled after the biological neural network in our brain, is the central piece of ChatGPT-4 and other modern AI systems. The AI neural network consists of layers of interconnect nodes. They behave like our neurons, which are the basic elements in our brain for information processing and transfer. Each node receives signals from adjacent nodes on the previous layers; adds these signals in a special way, or more precisely, performs a weighted summation of inputs; compares them with a threshold value; and if the threshold is exceeded, sends out the processed information to adjacent nodes on the subsequent layers, or in the context of a real neuron, starts firing. The information is then transformed and moved layer after layer until it reaches the end of the network. It is then outputted as the response to the user’s original request.   

The thing that the node in an AI network tries to model is actually the dendrite within a neuron, rather than the neuron itself. Dendrites are branched protoplasmic extensions in the neuron. For a picture and more information about the dendrite, see Wikipedia . One neuron may contain a number of dendrites. It is thus a network of dendrites in its own right, something actually more powerful than a node in the AI network.  

Furthermore, neuroscientists in Germany and Greece found in the year 2000 that even a human dendrite is more powerful than an AI node because the former can perform a logical calculation called XOR, which gives dendrites more control when processing the incoming information. Because of this capability, our brain can function more efficiently and use less resources, compared with the existing AI systems. Thus, there is room to further improve the AI neural network. It is possible to introduce an additional operation at each node, following the example of our brain; and this may help decrease the number of nodes in the AI network and thus reduce its complexity. This in turn has the potential to significantly lower the cost of developing an AI system. Considering that it may take months and millions of dollars to teach a large AI system nowadays, the improvement may be a big deal. To learn more about the discovery of the new computational power of our brain in 2000 and its significance, check out YouTube.  

As healthcare professionals who are devoted to the work on the human nervous system, we are eager to see our new knowledge of the brain help advance the technology again!