This project focuses on utilizing large quantities of calcium imaging neuron data to train CNN models. By combining state-of-the art neuron model and big data, the model was able to exhibit more complex behaviors from cells in the primary visual cortex.

The data were based on macaque monkey's neuron response given gray-scale image stimuli. These images were cropped samples from natural images. We then train CNN networks based on this data. The CNN networks recieve the images seen by the monkey as input, and are trained to predict the neuron responses as output.

Below is a visualization example for site m1s1. These images are generated using back-propagation on random images. In short, we fix a trained network for neuron prediction, and input random images to see the predicted response of neurons. We can thus use the model to get the gradient of the response (we use the negation of the response as loss) for each image pixel, and optimize the image to maximize the response. This basically allows us to create a estimation of each neuron's perceptive field.

Neuron Visualizations example

The effect of the big data is shown by this picture:

Correlation change example

This is the curve of change in correlation of 40 selected neurons. The correlation is measured by the real response values of the validation set images, versus the predicted response values of the same images (using the model). The x axis labels the number of samples used to train the models, and the training samples are in the original order in which they were presented to the monkey. Previous experiments usually don't collect data as much as 30k, but this graph shows the signifigance of data quantity to train effective models that can determine neurons' preference in different image stimuli.