BrainCog Cognitive Intelligence Engine

BrainCog includes integrated support for several standard neuromorphic and event-based datasets, making them available for brain-inspired AI research and simulation. Supported datasets include:

• DVSGesture: 11 hand gestures from 29 subjects under 3 illumination conditions (recorded using DVS128).
• DVSCIFAR10: 10,000 event streams converted from the frame-based CIFAR10 dataset.
• NCALTECH101: 100 object classes plus a background class, captured using an ATIS sensor on a motorized pan-tilt unit.
• ES-ImageNet: 1,000 categories converted from 1,300,000 ImageNet frame-based images using Omnidirectional Discrete Gradient (ODG).
• N-Omniglot: 1,623 categories of handwritten characters (20 samples per class) acquired via a DVS acquisition platform.

Dynamic Structure Development of Spiking Neural Networks (DSD-SNN) is designed for efficient and adaptive continual learning. It employs two primary mechanisms:

1. Dynamic Structure Adjustment: The network grows new neurons and prunes redundant ones to increase memory capacity while reducing computational overhead.
2. Knowledge Transfer: It overlaps shared structures to leverage knowledge acquired from previous tasks, allowing a single network to support multiple incremental tasks.

The implementation is validated against multiple Task-Incremental Learning (TIL) and Class-Incremental Learning (CIL) benchmarks.

When simulating human or macaque brain models, you can customize the simulation by modifying parameters within the main function of the .py files. Specifically, you can flexibly set:

• nsz: The number of neurons in each region.
• asz: The connectome power between regions.

You can customize the scale and connectivity of the brain simulation by modifying the main function in the .py files (e.g., brainSimMaq.py).

Specifically, you can flexibly set the following parameters:

• nsz: The number of neurons in each brain region.
• asz: The connectome power (connectivity strength) between regions.

To run the CRSNN example, ensure the following Python packages are installed in your environment:

• numpy
• scipy
• pytorch >= 1.7.0
• torchvision

The PL-SDQN (Potential Based Normalization for Spiking Deep Q Network) implementation is used to solve the spike feature information vanishing problem. To run the training process, execute the main.py script located in the sdqn directory.

Prerequisite: You must install the tianshou framework before running this model.

python ./sdqn/main.py

To train a deep spiking neural network (SNN) using the ANN-SNN conversion method, you must first prepare a PyTorch-based Artificial Neural Network (ANN). This involves replacing ReLU and MaxPooling layers in your PyTorch model with compatible alternatives to ensure the model can be successfully converted into an SNN for complex tasks.

Workflow:

1. Train the ANN: Run the initial training script to obtain a well-trained ANN.
2. Run SNN Inference: Use the converted script to perform inference using the spiking neural network.

# 1. Train the well-trained ANN
python CIFAR10_VGG16.py

# 2. Run the SNN inference process
python converted_CIFAR10.py

To get started with BrainCog, follow the official installation guide. The guide provides instructions for setting up the environment required for brain-inspired AI research and simulation.

Available versions:

• English Installation Guide
• Chinese Installation Guide

To use the SMAC environments with the PettingZoo multi-agent API, you must first install the pettingzoo package via pip.

pip install pettingzoo

To run the BAE-SNN (Brain-Inspired Affective Empathy Computational Model) example, ensure the following Python packages are installed in your environment:

• numpy
• scipy
• pytorch >= 1.7.0
• torchvision