Codes for a learning theory of binary simple BTSP

This repository contains the code for the Simple BTSP project.

Version Requirements

Python: 3.10
Torch: 2.1.0.dev20230317+cu117
Numba: 0.56.4
Scipy: 1.10.1

Execution Instructions

To validate different memory models, including BTSP, RP, and HFN, please execute the file named main_**.

Specific Model Usage

The BTSP Feedforward Model was utilized for Figures 2C-E, 3, 4, and 6.
The BTSP Feedback Model was employed for Figure 5.
Random Projections and the HFN with both continuous and binary weights were employed for Figures 3 and 5.
Functions for theoretical predictions can be found in theory_prediction_function.py.

Examples

To test HFN performance on non-orthogonal memory item tasks, please directly run the code in 'main_HFN_continuous_weights.py'; adjust the threshold (cdf_vth) to observe the impact of the threshold on performance. To test the performance on orthogonal memory items, simply replace the code lne 96 and 97 as used for BTSP models; To test the binarized HFN performance, please directly run the code in 'main_HFN_binary_weights.py'

We will release the complete versions of the codes after the publication of our study. Please cite the following reference for our work:

Reference