Pattern Storage and Retrieval in a Reduced Scale CA3-Inspired Neural Network
Open Access DepositedDownloadable Content
Understanding how biological neural circuits achieve efficient learning, stable memory storage, and pattern completion may provide valuable insights for designing more adaptive and energy-efficient artificial systems. This MSc thesis investigates the mechanisms for associative memory and pattern completion in a significantly compact model of the hippocampal CA3 region. Previous studies have shown that large, biologically plausible networks are able to perform these functions reliably, but it is still unknown if this is possible in networks with highly reduced sizes. In this work, we address three main research questions focused on (1) the ability of the reduced network to sustain stable attractor-like assemblies, (2) the contribution of different plasticity mechanisms at this scale network and (3) the impact of network reduction on memory capacity and recall performance. The reduced network contains 177 neurons, both excitatory pyramidal cells and inhibitory interneurons, and is simulated with the CARLsim 6 framework. The number of neurons was constrained using an evolutionary algorithm so that the resulting network shows periodic resting behavior and pattern completion capabilities with less than 20,0000 synapses for future implementation in a memristive-based neuromorphic platform. Neurons are modeled based on the Izhikevich formalism, and connections are probabilistic and type-specific. Learning is based on the symmetric Spike-Timing-Dependent Plasticity (STDP) rule for long-term plasticity and Short-Term Plasticity (STP) implemented using Tsodyks-Markram formalism . For a single pattern size of 20, stimulating half the assembly neurons, gives a maximum accuracy of 91.3%During the training process, structured input current activates specific neuronal groups repeatedly. Once the network is trained, it is then tested in partial-cue conditions to assess its performance in completing the pattern. We propose a composite performance measure that quantifies the activation of the target assembly, the suppression of the off-target assemblies, the specificity of the network's response, and the overall level of recall. Despite the substantial reduction in network size and connectivity, the model is able to recover key functional properties of the CA3 network. Network activity was evaluated within a 5 ms window centered on peak retrieval. Under 50% cue conditions, the network demonstrates strong pattern separation and selective retrieval, as indicated by low separation error (0.012) and precision error (0.027), showing that activity is largely confined to the target assembly. The recall error remains low (0.075), indicating that most of the assembly is successfully reactivated. At 75% cue, completion error decreases, precision improves, and recall error drops to zero. Importantly, separation error remains unchanged across both cue levels, demonstrating that improved retrieval is achieved without increased activation of non-target neurons. These results indicate that the reduced network supports robust and selective pattern completion, with retrieval quality scaling with cue strength. By showing that pattern storage and retrieval can be achieved at reduced scales, we provide insights into the structural and dynamical requirements for memory retrieval. Apart from the neuroscience implications, the ability to implement robust associative memory in small-scale networks is important for the design of neuromorphic systems where constraints on resources necessitate the design of compact yet effective systems.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.