DNNBrain: A Unifying Toolbox for Mapping Deep Neural Networks and Brains

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Cognitive Neuroscience

Background:

• Deep neural networks (DNNs) achieve human-level performance but lack transparency in their internal operations.
• Neuroscientists are interested in using DNNs to model biological systems and applying cognitive neuroscience to understand DNNs.
• Existing deep learning frameworks require advanced programming and mathematical skills, hindering interdisciplinary research.

Purpose of the Study:

• To introduce DNNBrain, a Python-based toolbox designed for cognitive neuroscientists.
• To facilitate the exploration of internal representations in both DNNs and biological brains.
• To bridge the gap between deep learning and cognitive neuroscience research.

Main Methods:

• DNNBrain integrates popular DNN software packages (e.g., PyTorch, TensorFlow) with established brain imaging tools.
• It offers both application programming interfaces (APIs) and command-line interfaces (CLIs).
• Key functionalities include extracting DNN activations, probing and visualizing DNN representations, and mapping DNNs to brain data.

Main Results:

• The toolbox provides a unified environment for cross-disciplinary research.
• It simplifies complex analyses, making DNN and brain representation mapping more accessible.
• Enables researchers to investigate DNNs as models of brain function and vice versa.

Conclusions:

• DNNBrain accelerates scientific discovery by simplifying the investigation of DNNs and their relationship to brain function.
• It lowers the barrier to entry for cognitive neuroscientists interested in deep learning.
• Facilitates a deeper understanding of the 'black box' of DNNs and biological neural systems.