Memory Encoding in the Hippocampus: Computational Models Illuminate CA3 Dynamics

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jan 31 2024

In a paper published in the journal Nature Communications, researchers explored the functional role of the hippocampal subfield cornu ammonis 3 (CA3), proposing its function as an auto-associative network for encoding memories.

The study unveiled dual input pathways from the entorhinal cortex (EC)and dentate gyrus (DG), highlighting processes of sparsification and decorrelation. The constructed CA3 model, resembling a Hopfield-like network, demonstrated the storage of dense and correlated and sparse and decor-related encodings.

Examining CA3 place cells in rats validated the model, showcasing pronounced tuning during theta phases with sparser activity. The study also suggested that incorporating complementary encodings in neural networks enhances multitasking learning, providing valuable insights for addressing complex tasks with computational advantages.

Related Work

Past research has underscored the importance of the hippocampus, particularly CA3, for forming episodic memories. CA3, viewed as an auto-associative network, possesses pyramidal cells with recurrent connections, facilitating pattern completion and retrieval from noisy cues. Sensory information reaches CA3 through the EC, with two pathways—direct synapses via the perforant path (PP) and an indirect route through the DG via mossy fibers (MF). DG sparsifies EC encodings, while MF maintains sparse, decor-related connectivity. PP, with dense connectivity, introduces correlated sensory information.

Integrated Approach Unravels CA3 Dynamics

In the experimental methodology, the researchers employed a multi-step approach to model the transformation of memories along hippocampal pathways. They used a binary autoencoder to process images from the Fashion-modified National Institute of Standards and Technology (MNIST) dataset. They trained a linear autoencoder with batch normalization and rectified linear unit (ReLU) nonlinearity.

The encoded representations were binarized to establish the desired density. Subsequently, binary feedforward networks were implemented from the EC to CA3, incorporating connectivity matrices and threshold values based on estimated biological values.

Researchers designed two key components to model the information flow within the hippocampus. First, they established a visualization pathway from CA3 to EC by training a feedforward network. They configured the network to map inputs and targets, enhancing our understanding of the dynamics between these hippocampal regions.

Concurrently, researchers actively developed a Hopfield-like model for CA3, enabling the system to store and retrieve patterns efficiently. This model played a crucial role in understanding the intricate mechanisms associated with information storage and recall within the CA3 region of the hippocampus. This dual approach provided insights into the complex interplay and computational roles of CA3 in processing and recalling diverse types of information.

The experimental data analysis encompassed considerations for calculating activity, information per spike, and sparsity correction for each neuron. The researchers applied these analyses to linear track and W-maze data, identifying place cells and decoding turn directions. Machine learning (ML) experiments used the MNIST dataset, training multilayer perceptrons for single-task and multitask learning. The loss functions included cross-entropy terms and additional terms to promote decorrelated representations. The training parameters and procedures varied for different tasks and activation functions.

Overall, the comprehensive methodology combined computational modeling, data analysis, and ML approaches to investigate the encoding and retrieval processes in the hippocampal subfield CA3, providing insights into memory-related functions and computational advantages.

Memory, Learning, Models

In the first study, researchers developed a computational model to investigate how memories are represented and transformed in the hippocampus, focusing on the pathways from the EC to the CA3 region. Using FashionMNIST images as sensory inputs, the model employed a sparse autoencoder in the EC to convert images into binary patterns. The projections from EC to CA3 resulted in distinct MF and lateral PP encoding patterns, with MF patterns being sparser and less correlated.

The model stored these patterns in a Hopfield-like network in CA3, demonstrating its ability to recall MF and PP encodings during retrieval. The study suggests that the model captures hetero-associative memory processes and exhibits unsupervised concept learning as more examples are stored, providing insights into how the hippocampus may transform and store memories.

In the second study, the researchers utilized a computational CA3 model to predict relationships between neural activity during theta oscillations and the encoding properties of hippocampal place cells. The study garnered support for the model's predictions by analyzing place cell data obtained from experiments conducted in W-maze and linear track environments.

Sparse theta phases were associated with increased information encoding about finer, example-like positions, while dense phases preferentially encoded coarser, concept-like positions. The findings underscore the role of theta oscillations in determining the scale of information representation in the hippocampus, shedding light on how neural dynamics contribute to encoding spatial and conceptual information.

In the third study, researchers explored the application of CA3-like complementary encodings in enhancing ML performance. Adapting a multilayer perceptron trained on MNIST handwritten digits, the model simultaneously performed digit classification and set identification tasks. Encoding loss functions, including DeCorr and HalfCorr, were introduced to manipulate correlations within the network's hidden layers, mirroring the CA3 model's correlated and decorrelated pathways.

Results indicated that encoding correlations influenced the network's performance in concept and example learning tasks. The HalfCorr approach, incorporating both correlated and decor-related encodings, outperformed others, highlighting the potential of leveraging diverse encoding scales for complex ML tasks.

Conclusion

In conclusion, the studies provide a holistic understanding of memory representation in the hippocampus, revealing insights into the neural dynamics during memory encoding and retrieval. The computational models effectively capture hetero-associative memory processes and showcase the hippocampus's ability for unsupervised concept learning.

Analysis of place cell data supports predictions about the role of theta oscillations in encoding spatial and conceptual information. Applying CA3-like encodings in ML further demonstrates the potential for improved task performance through diverse encoding scales. These findings contribute significantly to our knowledge of neural mechanisms, bridging the gap between biological processes and artificial intelligence applications.