ML Compression with HPE Labs

Model Compression Variational Autoencoder Model Distillation Specialised Hardware

I led an industry–academic study with Hewlett Packard Enterprise Labs on compressing high-rate sensor data. The work combined a variational autoencoder, boosted-tree distillation and specialised hardware implementation.

Results

• 12× data compression
• 24 ns inference latency
• 330 million operations per second
• 5× better energy efficiency than the baseline
• Systematic evaluation of accuracy, latency, throughput, memory and energy use

Publication

Co-author of “Memristive tabular variational autoencoder for compression of analog data in high energy physics”, arXiv:2602.15990. Submitted to Nature Communications.

Tech Stack

Python, PyTorch, variational autoencoders, boosted decision trees, model distillation, quantisation and hardware-aware ML.