DynVision

DynVision is a modular toolbox for building and evaluating recurrent convolutional neural networks (RCNNs) with biologically plausible temporal dynamics. The toolbox provides researchers with a framework to explore how details in the network architecture influence temporal dynamics and shape visual processing, while handling most of the overhead to achieve computational efficiently. The toolbox provides a collection of biologically-inspired components including:

realistic lateral recurrent connections
flexible skip and feedback connections
activity evolution governed by dynamical systems equations
unrolling of biological time with heterogenous time delays for different connection types

Documentation Structure¶

Our documentation is organized into four main categories:

Tutorials¶

Step-by-step guides for beginners to get started with DynVision.

Getting Started: First steps with DynVision
Basic Model Training: Train your first model
Custom Model Creation: Build your own neural network architecture

How-to Guides¶

Task-oriented guides for solving specific problems.

Installation: Detailed installation instructions
Custom Models: Define your own neural network architectures
Data Processing: Work with different datasets
Workflow Management: Use Snakemake for experiments
Model Testing: Evaluate model performance

Reference¶

Technical descriptions of DynVision's components.

Organization Overview: Structure of the toolbox
Model Components API: Core building blocks
Recurrence Types: Different recurrent connection implementations
Dynamics Solvers: ODE solvers for neural dynamics
Configuration Reference: Configuration file documentation

Explanation¶

Conceptual understanding of DynVision's approach.

Biological Plausibility: Alignment with neural systems
Temporal Dynamics: Understanding temporal properties
Design Philosophy: Core design principles

Citing DynVision¶

If you use DynVision in your research, please cite:

Gutzen, R. & Lindsay, G. (2025). Modeling Dynamical Vision with Biologically Plausible Recurrent Convolutional Networks. bioRxiv. doi:10.1101/2025.08.11.669756

@article{gutzen2025modelingdynamical,
  title   = {Modeling Dynamical Vision with Biologically Plausible
             Recurrent Convolutional Networks},
  author  = {Gutzen, Robin and Lindsay, Grace},
  year    = {2025},
  journal = {bioRxiv},
  doi     = {10.1101/2025.08.11.669756},
  url     = {https://doi.org/10.1101/2025.08.11.669756}
}

Contributing¶

DynVision is an open-source project, and we welcome contributions! See our Contributing Guide for information on how to get involved.

Getting Support¶

If you have questions or run into issues:

Search the GitHub Issues to see if someone has encountered the same problem. Open a new issue if you can't find a solution.
Reach out via Email.

License¶

DynVision is released under the MIT License. See the LICENSE file for more details.