Recurrent Neural Networks of Mediodorsal Thalamus and Prefrontal Cortex in Temporal Contexts

Model Framework

We develop recurrent neural networks with a thalamus-like component and synaptic plasticity rules to model the thalamocortical interactions in cognitive flexibility. We find that the MD component is able to extract context information by integrating context-relevant traces over trials and to suppress context-irrelevant neurons in the PFC. Incorporating the MD disjoints the contextual representations and enables efficient population coding in the PFC.

Get started with training and testing

• Train a default network with train_test_PFCMD_pytorch.py for the cognitive task in Rikhye et al. 2018

• Perform decoding analysis for context and rule with decoding_analysis.py

• Train a PFC-MD neural network on the Neurogym tasks with run_PFCMD.py in the CL_neurogym folder.

• The baseline continual learning methods, e.g., EWC and SI, are implemented in run_baselines.py.

• The model analysis is performed in run_analysis.py.

Dependencies

The code is tested in Python 3.6 and Pytorch.

Citation

If you use this code for your research, please cite our[paper:

@article{Zheng2024,
  title={Rapid Context Inference in a Thalamocortical Model Using Recurrent Neural Network},
  author={Wei-Long Zheng and Zhongxuan Wu and Ali Hummos and Guangyu Robert Yang and Michael M. Halassa},
  journal={Nature Communications},
  year={2024}
}

Related Projects:

PFC_MD_weights_stability: Code for the computational model to avoid catastrophic forgetting as in Rikhye, Gilra and Halassa, Nature Neuroscience 2018

ThalamusContextSwitchingCode: Code base to recreate figures from Rikhye et al. Nature Neuroscience paper.