Scaling Factor Grid Search

Functions for optimizing the scaling factor (lambda) used in rate-to-spike conversion. The scaling factor is crucial for maintaining task performance when converting from continuous rate dynamics to discrete spiking dynamics.

Main Functions

Grid Search Parameters

The main grid search function accepts:

• model_dir (str): Directory containing trained rate RNN model .mat files (default: ‘../models/go-nogo/P_rec_0.2_Taus_4.0_20.0’)

• task_name (str): Task type (‘go-nogo’, ‘xor’, or ‘mante’) (default: ‘go-nogo’)

• n_trials (int): Number of trials to evaluate each scaling factor (default: 100)

• scaling_factors (list): List of scaling factors to test (default: [20, 25, 30, …, 75])

Single Trial Evaluation

The evaluate_single_trial function tests a specific scaling factor:

• curr_full (str): Full path to model file

• scaling_factor (float): Scaling factor to test

• task_name (str): Name of the task to evaluate

• use_initial_weights (bool): Whether to use initial weights instead of trained weights

• down_sample (int): Temporal downsampling factor

Returns performance metrics for the given scaling factor.

Example Usage

from spiking import lambda_grid_search

# Basic grid search with default parameters
lambda_grid_search()

# Grid search with custom parameters
lambda_grid_search(
    model_dir='models/go-nogo',
    n_trials=50,
    scaling_factors=list(range(30, 81, 5)),
    task_name='go-nogo'
)

Optimization Process

The grid search follows these steps:

1. Load trained rate models from the specified directory

2. Generate test stimuli appropriate for the task type

3. Iterate through scaling factors in the specified range

4. Convert to spiking network for each scaling factor

5. Evaluate performance using task-specific metrics

6. Select optimal scaling factor based on performance

7. Save optimal value to model file for future use

Performance Metrics

Different metrics are used depending on the task:

Go-NoGo Task:

• Response accuracy for Go trials

• Response inhibition accuracy for NoGo trials

• Combined accuracy score

XOR Task:

• Output accuracy for all input combinations

• Temporal precision of responses

Mante Task:

• Context-dependent decision accuracy

• Sensory integration performance

Output Files

The function modifies each input .mat file to include:

• opt_scaling_factor: The optimal scaling factor found

• all_perfs: Performance scores for all tested scaling factors

• scaling_factors: List of all scaling factors that were tested

Output

The function outputs:

• Progress updates during optimization

• Performance scores for each scaling factor tested

• Optimal scaling factor and its performance

• Updated model file with optimal scaling factor saved