Companion post to: “Example Based Hebbian Learning may be sufficient to support Human Intelligence” on Biorxiv.
With the tremendous success of deep networks trained using backpropagation, it is natural to think that the brain might learn in a similar way. My guess is that backprop is actually much better at producing intelligence than the brain, and that brain learning is supported by much simpler mechanisms. We don’t go from Zero to super smart in hours, even for narrow tasks, as does AlphaZero. We spend most of our first 20 years slowly layering into our brains the distilled intelligence of human history, and now and then we might have a unique new idea. Backprop actually generates new intelligence very efficiently. It can discover and manipulates the huge dimensional manifolds or state spaces that describe games like go, and finds optimal mappings from input to output through these spaces with amazing speed. So what might the brain do if not backprop?
We know that STDP happens and that it can implement local Hebbian learning rules. We also know that classical conditioning happens, and I think the observation of classical conditioning provides an important clue. Classical conditioning says that if A causes B, and A is juxtaposed with some other input X with appropriate timing, then the animal learns to causally link X with B. This suggests that neural pathways exist between X and B, but were initially not configured with synaptic weights that would support a causal link between them. Before conditioning, B is caused by A, and X produces some unrelated activity. If Hebbian learning of the STDP variety is active during conditioning at the synapses on the path from X to B, then those synapses whose activity from X happened to arrive at B with appropriate timing to cause B would be strengthened. This will tend to produce connections from X to B that would be causal in the future, thus implementing classical conditioning.
We argue in this paper that this form of Hebbian learning is a good candidate for a mechanism for classical conditioning. But if this is the case, then what other learning phenomena might be explained by Hebbian learning that is not simply reinforcing already causal connections, but creating new causal connections from example behavior? If such a Hebbian mechanism exists, then one can generalize and postulate that any pattern of neural activity that is produced by some generator of example behavior can be attached to any other pattern that precedes it with appropriate timing. In the paper we call this type of learning Example Based Hebbian (EBH) learning, and show with simple examples that this class of learning mechanism can support procedural, episodic, and semantic memory.
We further speculate that this EBH learning mechanism can support human intelligence.
We all have a mental workspace within which we can think about things other than the immediate physical present. Some might call it our conscious thought process, or our imagination, or our internal world model. Whatever it is called, it differentiates us quite dramatically from most other animals. We speculate that this process allows for extremely flexible generation of example behaviors and associations that can be set into the brain using a simple EBH mechanism. These new associations and behaviors can then be used by the conscious thought process itself to generate yet more advanced behavior, and our intelligence builds. With backprop, or mechanisms that do the equivalent of backprop, the key is in the learning algorithm itself. You just need a cost function and some example data, and the learning algorithm reconfigures the network to optimize the cost. With EBH learning, the learning algorithm is nearly trivial gluing together of associations, and the key is in the generation of good examples by conscious thought. Which sounds more like how humans learn?