<p>In the field of neuroscience, it has been confirmed that the hippocampal CA1 plays a crucial role in spatial navigation. Establishing an appropriate deep learning architecture to replicate this process can not only more accurately predict behavior but also better understand the working mechanisms of CA1. This study utilizes an open-source dataset to predict rat immediate behavior by constructing a hybrid network capable of processing spatial-temporal information and dynamically activating over input. Specifically, the KA-AttLSTMnet architecture is proposed to replicate the CA1’s navigation function. First, this study employs spike-related methods to encode CA1 activity and represents a total of eight open-field behavioral states under allocentric/egocentric strategies. Then, spike encoding and behavioral states are used as inputs and outputs, respectively, to establish eight different immediate behavior prediction models. Finally, in terms of model structure, a self-attention mechanism architecture based on recurrent neural network (RNN) is built, and Kolmogorov-Arnold network (KAN) is employed to further dynamically adjust the architecture, which already possesses spatial-temporal processing capabilities, to enhance the extraction of CA1 neural activity information and improve the prediction of the rat’s immediate behavior. By comparing the prediction results of the eight behavioral states, only two behaviors, direction of turn <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({{{\varvec{b}}}_{{\varvec{t}}}}^{{\varvec{i}}}\)</EquationSource> </InlineEquation> and speed <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\({{b}_{s}}^{i}\)</EquationSource> </InlineEquation>, achieved relatively stable predictions, indicating that the CA1 internal circuit is more inclined to fully reflect an egocentric strategy rather than an allocentric one. Ablation experiments demonstrated that the unrolled-LSTM network is more effective in processing spike encoding over time and organizing it over space. Additionally, the mean squared error (MSE) for predicting <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\({{{\varvec{b}}}_{{\varvec{t}}}}^{{\varvec{i}}}\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\({{b}_{s}}^{i}\)</EquationSource> </InlineEquation> decreased from 0.4203/0.0435 to 0.3687/0.0150, and eventually to 0.3255/0.0110. This reduction highlights the positive impact of the multi-head self-attention mechanism in RNN (AttLSTMnet) for extracting contextual information, as well as the dynamic regulation capability of neurons in the KAN (KA-AttLSTMnet), which differs from the traditional weight-activation function mechanism, both of which contribute significantly to improving the immediate behavior prediction model.</p>

KA-AttLSTMnet: a Kolmogorov-Arnold attentional architecture for egocentric navigation prediction from hippocampal CA1 spikes

期刊论文
发表日期：

2026年1月8日

查看全文

Xiaolong Wu,
Jianhong Yang,
Zhanhong Du

摘要

In the field of neuroscience, it has been confirmed that the hippocampal CA1 plays a crucial role in spatial navigation. Establishing an appropriate deep learning architecture to replicate this process can not only more accurately predict behavior but also better understand the working mechanisms of CA1. This study utilizes an open-source dataset to predict rat immediate behavior by constructing a hybrid network capable of processing spatial-temporal information and dynamically activating over input. Specifically, the KA-AttLSTMnet architecture is proposed to replicate the CA1’s navigation function. First, this study employs spike-related methods to encode CA1 activity and represents a total of eight open-field behavioral states under allocentric/egocentric strategies. Then, spike encoding and behavioral states are used as inputs and outputs, respectively, to establish eight different immediate behavior prediction models. Finally, in terms of model structure, a self-attention mechanism architecture based on recurrent neural network (RNN) is built, and Kolmogorov-Arnold network (KAN) is employed to further dynamically adjust the architecture, which already possesses spatial-temporal processing capabilities, to enhance the extraction of CA1 neural activity information and improve the prediction of the rat’s immediate behavior. By comparing the prediction results of the eight behavioral states, only two behaviors, direction of turn \({{{\varvec{b}}}_{{\varvec{t}}}}^{{\varvec{i}}}\) and speed \({{b}_{s}}^{i}\) , achieved relatively stable predictions, indicating that the CA1 internal circuit is more inclined to fully reflect an egocentric strategy rather than an allocentric one. Ablation experiments demonstrated that the unrolled-LSTM network is more effective in processing spike encoding over time and organizing it over space. Additionally, the mean squared error (MSE) for predicting \({{{\varvec{b}}}_{{\varvec{t}}}}^{{\varvec{i}}}\) and \({{b}_{s}}^{i}\) decreased from 0.4203/0.0435 to 0.3687/0.0150, and eventually to 0.3255/0.0110. This reduction highlights the positive impact of the multi-head self-attention mechanism in RNN (AttLSTMnet) for extracting contextual information, as well as the dynamic regulation capability of neurons in the KAN (KA-AttLSTMnet), which differs from the traditional weight-activation function mechanism, both of which contribute significantly to improving the immediate behavior prediction model.