braincoder: (Encoding &) Decoding Neural Data

Gilles de Hollander

braincoder

braincoder

Key features:

• Massive computationa optimisation thanks to Tensorflow and GPUs
• Ease-of-use
• All standard models implemented
• Inversion of encoding models

1. Encoding Models: Mapping Stimulus → BOLD

2. Linear Encoding Models

3. Building the Likelihood

• Add Gaussian noise to deterministic models:

• Covariance matrix ():

  • Regularized estimate (shrinkage + neural population overlap).
  • Accounts for voxel-to-voxel noise correlations.

Key:

• Enables Bayesian inversion (stimulus decoding).

4. From Neural Responses → Stimulus Features

4. From Neural Responses → Stimulus Features

5. Decoding Noisy Neural Data

Key Takeaways

• Encoding: Map stimuli to BOLD (linear/non-linear models).
• Decoding: Invert models using Bayesian inference + noise modeling.
• Tools:
  • braincoder can do all of this and leverages TensorFlow for fast, GPU-accelerated fitting.

Your Turn:

• Try fitting a PRF model to your own data!

Example code for PRF fit

from braincoder.models import GaussianPRF2DWithHRF
from braincoder.hrf import SPMHRFModel
from braincoder.optimize import ParameterFitter

# Set up model, including HRF
hrf_model = SPMHRFModel(tr=1.7)
model = GaussianPRF2DWithHRF(grid_coordinates=grid_coordinates, hrf_model=hrf_model)

# Set up fitter
fitter = ParameterFitter(data=v1_ts, model=model, paradigm=stimulus)

# Define grid search parameters
mu_x = np.linspace(-3, 3, 20, dtype=np.float32)
mu_y = np.linspace(-3, 3, 20, dtype=np.float32)
sigma = np.linspace(0.1, 5, 20, dtype=np.float32)
baselines = [0.0]
amplitudes = [1.0]

# Do grid search using correlation cost (so baseline and amplitude do not matter)
grid_pars = fitter.fit_grid(mu_x, mu_y, sigma, baselines, amplitudes, use_correlation_cost=True)

 # Refine baseline and amplitude using OLS
grid_pars = fitter.refine_baseline_and_amplitude(grid_pars)
gd_pars = fitter.fit(init_pars=grid_pars)

Assigment 4: Decoding visual stimuli

Open notebooks/4_decode.ipynb.