A Computational Model for Storing Memories in the Synaptic Structures of the Brain

Abstract

Spike-timing dependent plasticity (STDP) is widely accepted as a mechanism through which the brain can learn information from different stimuli(1, 2). Basing synaptic changes on the timing between presynaptic and postsynaptic spikes enhances contributing edges within a network(3, 4). While STDP rules control the evolution of networks, most research focuses on spiking rates or specific activation paths when evaluating learned information(5–7). However, since STDP augments structural weights, synapses may also contain embedded information. While imaging studies demonstrate physical changes to synapses due to STDP, these changes have not been interrogated based on their embedding capacity of a stimulus(8–12). Here, we show that networks with biological features and STDP rules can embed information on their stimulus into their synaptic weights. We use a k-nearest neighbor algorithm on the synaptic weights of thousands of independent networks to identify their stimulus with high accuracy based on local neighborhoods, demonstrating that the network structure can store stimulus information. While spike rates and timings remain useful, structural embed-dings represent a new way to integrate information within a biological network. Our results demonstrate that there may be value in observing these changes directly. Beyond computational applications for monitoring these structural changes, this analysis may also inform investigation into neuroscience. Research is underway on the potential of astrocytes to integrate synapses in the brain and communicate that information elsewhere(13–15). In addition, observations of these synaptic embeddings may lead to novel therapies for memory disorders that are difficult to explain with current paradigms, such as transient epileptic amnesia.Significance StatementLearning in the brain is often achieved via spike-timing dependent plasticity changing the structure of synapses to augment the strength between neurons. Typically, these changes contribute to other behaviors in the network, such as spiking rates or spike timings. However, observing these changes themselves may be fruitful for interrogating the learning capability of networks in the brain. Using a computational model, we demonstrate that the synaptic weights contain an embedding of the stimulus after a certain amount of recurrent activity occurs. It is possible that networks in the brain embed information in a similar way and that external readers, such as astrocytes, can interrogate, integrate, and transport this synaptic weight information to process stimuli.