Here, we provide a brief overview of our neural predictive model.
For our competition, we have trained two baselines:
- Our state-of-the-art CNN model
- a simple linear-nonlinear (LN) model
Have a look our demo notebook as an introduction as to how our CNN model works.
We also show how to train a model on multiple datasets at once, compute its performance, and show the retinotopy of the V1 neurons by relating the learned RF position to the anatomical coordinates with V1: Generalization demo.
We recommend to use a machine with a CUDA compatible GPU. On a CPU, the model training and evaluation is possible, too, but quite slow. To use CPU only, set device=cpu
instead of device=cuda
in the example notebooks.
- This Video goes into the details of how our model is built (prepared for ICLR 2021).
- Generalization in data-driven models of primary visual cortex.
- the ICLR publication corresponding to the video above, which
- Behavioral state tunes mouse vision to ethological features through pupil dilation
- Uses a similar architecture as the CNN above, but also utilizes the behavioral variables to train the model. The same techniques are used in our baseline models for the Sensorium+ track.