Algonauts2023

 The new Algonauts challenge is out http://algonauts.csail.mit.edu/challenge.html and it's about encoding models!

First idea: use activities in brain areas to predict activities in other brain regions. How much variation does this explain?

Second idea: use a bunch of different pre-trained deep learning models, including ResNet and Visual Transformer and a linearizing approach to predict the activities of different ROI's.

Third idea: combine the predictions from different sources in some statistically optimal way. I have no idea how to do this, but seems interesting. 

Different models will predict variance along different dimensions of the input space. They will learn different transformations, different dimensions in the target space. How can we combine these directions? Perhaps we can just sum these vectors (weighted by learned coefficients) together as the scale doesn't matter. The metric for evaluation is correlation after all. 

Let the work begin.