Project repository for Fall 2023.

Additional project updates and ideas are tracked here in this Google document.

Need to debug/figure out why model(s') performance is/are drastically different from notebooks 4 and 6 & 7.

Results are located in sandbox4 notebook. At the moment, not all tasks assigned from Mon., 14 Aug 2023 have been completed. So far:

• [] Add option to store gradient norm of each layer, stored separately
• Change linear layers to: CNN + 1 linear layer
• Make deep model (5 layers), and train it to perfection (up to 99% or higher train accuracy)
• Save the model (we’ll call this the “ground model”) (if time, create 5 ground models)
• [] Then, create 10 models per noise level (pick 10 noise levels, between totally destroyed and basically no impact) (also loop through which layer)→ turns into 500 models. Make them noisy, measure all the things above (robustness, generalization.1, try generalization.2)
• [] Can experiment with gradcam (interesting but not most important)
• Try training with 32 all the way (in conv layers) - see if model can still be 99% good
• Use a smaller model, smallest non-trivial model
• Reduce number of linear layers
• Start profiling (draw on piece of paper)
• [] Look for number of weights in each model
• [] Get model training up to 100%

• Add option to store gradient norm of each layer, stored separately
• Create table, row -> model, col -> specs (grad norm, layerwise norm, specify train/test accuracy), list number of tunable parameters for each model.
• Add norms of total and/or per layer parameters to the table.
• [] GradCam (wishlist or next step)

Results are located in sandbox4 notebook. At the moment, not all tasks assigned from last week have been completed. So far:

• Add option to store gradient norm of each layer, stored separately
• Change linear layers to: CNN + 1 linear layer
• Make deep model (5 layers), and train it to perfection (up to 99% or higher train accuracy)
• Save the model (we’ll call this the “ground model”) (if time, create 5 ground models)
• Then, create 10 models per noise level (pick 10 noise levels, between totally destroyed and basically no impact) (also loop through which layer)→ turns into 500 models. Make them noisy, measure all the things above (robustness, generalization.1, try generalization.2)
• Can experiment with gradcam (interesting but not most important)

Current results are located in the sandbox3 notebook. The results mainly include a function whose input is a model and it returns a dictionary of gradients for both the weights and the bias(es). The dataset used was the MNIST dataset. The model used was a simple 5-layer neural network.