CONCEPT
Synaptic Plasticity vs. Weight Update
The foundational parallel and the foundational divergence of the AI age: both biological memory and artificial learning work by changing the strengths of connections among units, but the mechanism, locality, sample-efficiency, and biological stakes of the change are so different that sharing the vocabulary of “learning” papers over a chasm that runs through every serious question about what machines can know.
The vocabulary of artificial intelligence is, at its core, borrowed from neuroscience. The artificial “neuron” sums its inputs; the artificial “synapse” is a weight, a number scaling one unit’s influence on another; “learning” is a rule that nudges those weights so the system’s behavior improves. Every layer of this vocabulary traces back to
Eric Kandel’s biology: the discovery that a memory is a change in the strength of a synaptic connection, and that the pattern of those altered connection-strengths is what the animal has learned. The artificial network and the biological brain are thus genuine cousins at the level of abstract description: both are connectionist systems that store what they know in patterns of connection-strengths rather than in a library of explicit rules, and both can generalize because the connection pattern