Projects – Compression for ML applications

What compression schemes can best support learning over communication constrained networks? Compression forms a core component of efficient communication schemes, where the traditional goal is to represent data using a limited number of bits as precisely as possible, optimizing a rate-distortion trade-off. Instead, in this line of work, we design compression schemes tailored to machine learning applications: our quantization schemes are not optimized for reconstruction accuracy, but to support a specific machine learning task, such as classification or model training.