Directions
Efficient ML Inference
High-throughput, low-latency, and low-cost inference systems and techniques for ML models (traditional DNNs and modern LLMs).
Efficient ML Training
Efficient, high-throughput, fair, and cost-effective training and fine-tuning systems for ML models (traditional DNNs and modern LLMs).
Compound AI Systems
Designing and optimizing systems that tackle AI tasks by combining multiple interacting components, including retrievers (Retrieval Augmented Generation), multiple calls to models, or external tools (Agents).
ML for Systems
Applying ML (reinforcement learning, Bayesian optimizations, LLMs, etc.) to key systems decisions like caching, resource management, anomaly detection, and network protocol design.
Novel Model Architectures
Full-stack innovations on State Space Models (e.g., Mamba), and more generally, models with linear Attention/RNN layers.
Edge AI Systems
Video analytics, video conferencing, and video streaming.
Emerging Paradigms
Multimodal models, diffusion models, MoEs, inference-time scaling, etc.
Hardware Design for ML
Tools for exploring the hardware design space, silicon cost aware hardware design for ML workloads, hardware for serving large models.
LLM Safety
Building systems and datasets for improving LLM robustness, misuse prevention, and transparency.
Sequence Modeling
Developing model architectures for sequence modeling and optimizing the processing of long context windows.
Intelligent Resource Management
Scheduling, multi tenant GPU sharing, and learned cache eviction for ML workloads.
Security and Privacy
Privacy preservation for ML to prevent leaking and misusing sensitive information, efficient ML models for detecting threats.
Data Systems
Leveraging ML and optimization techniques to build novel designs for data storage layouts, database indexes, and end-to-end data systems.
