Uninformative memories will prevail: the storage of correlated representations and its consequences
Autoassociative networks were proposed in the 80’s as simplified models of memory function in the brain, using recurrent connectivity with Hebbian plasticity to store patterns of neural activity that can be later recalled. This type of computation has been suggested to take place in the CA3 region of the hippocampus and at several levels in the cortex. One of the weaknesses of these models is their apparent inability to store correlated patterns of activity. We show, however, that a small and biologically plausible modification in the “learning rule”, (associating to each neuron a plasticity threshold that reflects its popularity) enables the network to handle correlations. We study the stability properties of the resulting memories (in terms of their resistance to the damage of neurons or synapses), finding a novel property of autoassociative networks: not all memories are equally robust, and the most informative are also the most sensitive to damage. We relate these results to category-specific effects insemantic memory patients, where concepts related to “non-living things” are usually more resistant to brain damage than those related to “living things,” a phenomenon suspected to be rooted in the correlation between representations of concepts in the cortex.
Emilio Kropff and Alessandro Treves, 2007
I thought this article was interesting. It presents a theoretical analysis of neural networks, a common flaw in their implementation, and a biologically plausible solution. Their findings are relevant to HTM systems.