THINGSvision

THINGSvision extracts activations from deep neural networks (DNNs) to enable quantitative comparison of DNN representations with biological and behavioral data in cognitive science and computational neuroscience.

Key Features:

• Layer-wise activation extraction: Extracts activations across multiple layers of DNN architectures for use in downstream analyses.
• Model support: Supports both pretrained and randomly-initialized neural network architectures.
• Representational similarity analysis (RSA) integration: Computes RSA to compare representational structures between DNN activations and other datasets.
• Multimodal comparison: Relates DNN activations directly to functional MRI and behavioral datasets.
• Custom image dataset feature extraction: Extracts features from user-provided image datasets for model–brain–behavior studies.
• Reproducibility facilitation: Provides standardized extraction and analysis outputs to support reproducible research.

Scientific Applications:

• Model–brain comparison: Compare DNN representational geometries to functional MRI measurements of visual systems.
• Model–behavior comparison: Relate DNN activation patterns to behavioral datasets and responses.
• Assessing DNN validity for vision: Evaluate how DNNs that perform object classification relate to signals recorded from biological visual systems.
• Feature provision for cognitive neuroscience: Provide extracted features from image datasets for experimental and computational studies in cognitive science and computational neuroscience.

Methodology:

Extract activations from multiple layers of pretrained and randomly-initialized DNN architectures and compute representational similarity analysis (RSA) to compare DNN activation patterns with functional MRI and behavioral datasets, including feature extraction from custom image datasets.