Pretraining of Embodied Recurrent Networks Bridges the Gap Between Artificial and Cortical Neural Activities

This repository accompanies the paper "Pretraining of Embodied Recurrent Networks Bridges the Gap Between Artificial and Cortical Neural Activities". It trains RNNs in a biomechanical simulation environment (MotorNet) through a developmental pretraining pipeline and evaluates alignment between model hidden activity and macaque motor cortex recordings using CCA and DSA.

Repository Structure

Pretrained_RNN/
├── model/                  # RNNCell (Dale's law, sparse connectivity) & builders
├── tasks/                  # Motor tasks: grid reach, center-out, double reach, RTT
├── training/               # Training scripts & hyperparameters
├── scripts/                # Data preprocessing (firing rates, CCA/AUC)
├── analysis/               # Neural similarity analysis & statistics
├── figures/                # Paper figure generation (Fig 3–7)
├── Data/                   # Neural data & precomputed results
│   ├── CO/                 # Center-out task data
│   ├── DR/                 # Double-reach task data
│   └── RTT/                # Random-target task data
├── Results/                # Output directory
├── requirements.txt
└── setup.py

Installation

pip install -e .

Or install dependencies manually:

pip install -r requirements.txt

Usage

Training (progressive stages):

python -m training.pretrain_grid   # Stage 1: grid reach pretraining
python -m training.train_co        # Stage 2: center-out fine-tuning
python -m training.train_dr        # Stage 3: double-reach training
python -m training.train_rtt       # Stage 4: random-target sequence training

Data preprocessing & analysis:

python -m scripts.pre_data         # Generate model firing rates
python -m scripts.pre_CO_data      # Compute CCA/AUC
python -m analysis.eval_BasicModel # Behavioral evaluation
python -m analysis.Compute_CO_DSA  # DSA computation
python -m analysis.Melting         # Cross-model CCA analysis

Figures:

python -m figures.fig3   # Fig 3–7 generation

Model

RNNCell (model/rnn.py) features Dale's law (80% excitatory / 20% inhibitory), sparse connectivity (recurrent density 0.8), spectral radius normalization (1.5), a separate task-input MLP, and configurable sensory feedback (proprioceptive, visual, or combined). Time constants: dt=10ms, tau=50ms.

Data

• Data/CO/, Data/DR/, Data/RTT/ — model firing rates, CCA/AUC/DSA results, and trajectory data for each task.

• Monkey C (fr_MC_*.npy) and Monkey M (fr_MM_*.npy) motor cortex recordings are included for neural comparison.

• Pretrained weights are in model/.