Our long-term goal is to make AI more responsible–accessible, sustainable, and righteous.

As the first step, we focus on various facets of Efficient ML, which helps us make AI equally accessible to everyone.

In particular, we work on three dimensions:

• Inference. We develop fast, low-resource methods to serve massive multimodal AI, e.g.,
  • Model Compression
  • Parallel Decoding
  • Batch Scheduling
• Training. We resolve the compute / memory bottlenecks for training large-scale models, e.g.,
  • Meta-Learning
  • Model Merging
  • Parameter-Efficient Fine-Tuning
• Data Dimension. We design algorithms to handle data with extremely high dimensionality, e.g.,
  • Data Compression
  • High-res Video Processing
  • Time-Series Forecasting

See our papers for more.