Neural Network-Based Decoding Input Stimulus Data Based on Recurrent Neural Network Neural Activity Pattern

Area of Science:

• Computational neuroscience
• Artificial intelligence

Background:

• Artificial neural networks (ANNs) store information through dynamic excitation patterns.
• Recurrent neural networks (RNNs) are capable of processing temporal sequences and maintaining information over time.

Purpose of the Study:

• To assess the feasibility of recovering information processed by an ANN by examining its neural activity patterns.
• To develop a method for decoding stored information from neural activity.
• To identify the minimal neural components essential for representing stimuli.

Main Methods:

• Utilized a simple recurrent neural network (RNN) model.
• Employed an advanced delayed match-to-sample task with variable pause durations.
• Developed a neural network-based decoding method to analyze excitation patterns.
• Identified minimal neuronal subsets containing stimulus information.

Main Results:

• The RNN successfully formed dynamic excitation patterns to store stimulus data.
• Invariant representations of stimuli were detectable within a specific time window (3-6 clock cycles).
• The proposed decoding method achieved 100% efficiency in recognizing received stimuli.
• A minimal subset of neurons was identified as sufficient for comprehensive stimulus information.

Conclusions:

• Information processed by ANNs can be recovered by analyzing neural activity patterns.
• The developed decoding method offers a highly effective way to decode neural representations.
• Understanding these minimal neuronal subsets advances insights into neural information processing and storage.