Frugal and efficient AI

Deep models often need to run far from the data center: on phones, on cameras, on embedded devices. Frugal AI addresses this end-to-end — from theory (what is the right notion of compressibility?) to algorithms (pruning, quantization, low-rank decomposition, knowledge distillation) to practical pipelines that deliver real efficiency without losing task performance.

Recent work in the group has explored task-aware and data-free compression, structured sparsity, and information-theoretic measures of redundancy in neural networks.