The Sampling Brain.

Consider the Hopfield network [3], a “neural network,” with symmetrical connections between binary neural “units.” Hopfield showed how such a network could learn: patterns were “imposed” on the network, and connections modified by local Hebbian learning. Remarkably, the network could “fill in” patterns from fragments, providing a form of “content-addressable memory.” Hopfield showed, too, that the ‘free-running’ of such a network minimized an “energy function” across the entire network, measuring the coherence of the pattern with respect to the connection weights (roughly, coherence involves positive weights between units with the same value; negative weights between units with different values). The behavior of the network as it falls into a stable pattern can be viewed as falling into an attractor basin---just as the dynamics of many physical systems can be modelled as descending in an energy landscape.