Thoughts after 12 hours of working with Algonauts2023 data

 I love competitions. I always want to win. It drives me. I thrive on competitions. I want to win Algonauts2023. I am doing it for the learning. I watched the first 40 mins of an absolutely amazing video on using deep learning models in fMRI https://www.youtube.com/watch?v=jobQmEJpbhY It's on an approach called voxel-wise encoding. Amazing visualizations! Images of brain neural patterns are now burned into my brain. There must be an area in my brain that reacts strongly to these images (it's the cortex flattened on a 2D surface, flatmap, see a library called pycortex).

So this is amazing stuff. You take a neural network trained on a task and pass some data through it and get some encodings out of this data and then use a linear model on these intermediate features to predict brain activity. Okay, I made it sound simple. In reality deep learning models are pretty complicated. Which activity patterns in which layers do you use to predict the neural activity? Or do you use multiple and then use something like bagging to combine models? Or should you use Bayesian model averaging? And there's loads of pre-trained models? Do you grab a pose estimator or a Visual Transformer or ResNet trained to recognize emotions in faces? OMG, my brain is exploding! Here's what I have so far https://github.com/mariakesa/ImagesToVoxels I fully intend to work on this challenge for the whole extent. Messing around with the data and the sample code has already taught me loads of practical things. Somebody once said, "Always have a challenge!". I want to get a deep learning job and this stuff is difficult, but really useful. I'm making a Streamlit visualization site that I can link to as an example project in job applications. There's really awesome free books online www.d2l.ai and www.learnpytorch.io Great stuff! I intend to use the opportunity to learn a lot about the brain and then try out my hand on NeuroPixel probe data next. There's even NeuroPixel probe data for humans published with a Nature paper. This stuff will blow your mind! Also, I think I know where I want to do a PhD. Caltech, the Fear Lab.