Modular inhibitory coding in binary networks

We characterized properties of class of binary models where, as observed in biological networks, excitatory neurons are structurally and functionally separated from inhibitory units. We investigate the respective roles the two populations play in memory storage.

The network is composed of separated excitatory and inhibitory layer. New patterns, represented as activation and inactivation of binary units in excitatory layer, are stored in the network through recruitment and training of inhibitory units that are grouped into individual modules and interact with excitatory layer. At the same time, the inhibitory modules compete for activation based on the signal magnitude they receive from the excitatory layer.

We show that inhibitory layer plays a critical role in memory storage and management, and that capacity of the network scales proportionally to number of inhibitory neurons. Further, we demonstrate that performance of the network is only gradually diminished when excitatory-to-excitatory (E-E) connections are removed but critically depends on inhibitory-to-excitatory (I-E) connections. We further show advantages of so designed coding scheme in terms of memory capacity, its expansion with progressive storage of new memories as well as network behavior for large memory loading.

These results are in line with new experimental work showing that inhibitory interneurons are playing critical role in memory storage and recall in the brain networks and may also address why generally excitatory networks exhibit sparser reciprocal connectivity as compared to connections to/from inhibitory units.