THINGSvision: A Python Toolbox for Streamlining the Extraction of Activations From Deep Neural Networks.

Area of Science:

• Computational neuroscience
• Cognitive science
• Machine learning

Background:

• Deep neural network (DNN) models show near-human performance in object classification and predict biological visual system signals.
• Understanding DNNs requires extracting layer activations, which can be complex for non-experts.
• Existing methods for DNN activation extraction are often unwieldy and error-prone.

Purpose of the Study:

• To introduce THINGSvision, a Python module designed to simplify the extraction of DNN layer activations.
• To provide a unified tool for accessing activations from various neural network architectures.
• To facilitate the use of DNNs in cognitive neuroscience research by lowering the computational barrier.

Main Methods:

• Development of the THINGSvision Python module for streamlined DNN activation extraction.
• Utilizing a wide range of pretrained and randomly-initialized neural network architectures.
• Integration of representational similarity analysis (RSA) for relating DNNs to empirical data.

Main Results:

• THINGsvision offers a user-friendly interface for extracting layer activations.
• The module successfully extracts features from diverse DNNs for custom image datasets.
• Demonstrated utility by linking DNN activations to functional MRI and behavioral data via RSA.

Conclusions:

• THINGsvision enhances the ability of researchers to relate DNNs, brain activity, and behavior.
• The module improves the reproducibility of findings in cognitive science and computational neuroscience.
• Facilitates interdisciplinary research by simplifying complex computational tasks.