My research applied geometric models to deep learning to generate new views of data to improve performance with limited labels. To achieve this, I would use techniques like **sparse coding**, **manifold learning**, and **self-supervised learning**. I am interested in building large foundation models with sparse intermediary layers (i.e., only a few elements of each activation non-zero) for interpretability, robustness, and compute efficiency. One approach I have developed uses **variational inference** to reach similar performance as unrolled optimization routines, like [FISTA](http://www.stat.cmu.edu/~ryantibs/convexopt/lectures/prox-grad.pdf), with a single neural network forward pass.
0 commit comments