<aside> 💡 ML models have grown TOO BIG—let’s fix this.

</aside>

Big models work surprisingly well. But are also surprisingly bad: (1) Slow inference, (2) Energy- and CO2-heavy training, (3) Requires too much data, (4) Contains many undiscoverable biases, …

EffL aims to resolve this “big model problem” through:

• Network Compression We make neural networks smaller (e.g., pruning, quantization, distillation, …)
• Data Compression We make the data smaller (e.g., Image/Video Compression, Dataset distillation, …)
• Training Acceleration We make the model training faster (e.g., better optimizers/preprocessing, model merging, …)

We hope to gain and utilize expertise on theory, algorithm, and system to give full-stack solutions.