Skyplane: Optimizing Transfer Cost and Throughput Using Cloud-Aware Overlays
NSDI 2023 USENIX Symposium on Networked Systems Design and Implementation
Skyplane accelerates wide-area transfers in the cloud by 5x via overlay routing and parallelism. Our open-source project moves TBs of data reliably and efficiently (>20Gbps) at up to 4x lower costs.
POET: Training Neural Networks on Tiny Devices with Integrated Rematerialization and Paging
ICML 2022 International Conference on Machine Learning
We present POET, an algorithm to enable training large neural networks (like BERT) on memory-scarce battery-operated edge devices.
Learning to Design Accurate Deep Learning Accelerators with Inaccurate Multipliers
DATE 2021 Design, Automation and Test in Europe
We improve the power-efficiency of the TPU by 4-6% by synthesizing a novel low-power approximate version using learning-augmented search.
Contrastive Code Representation Learning
EMNLP 2021 Empirical Methods in Natural Language Processing
Learning to represent software functionality for automated software engineering tasks like type inference, clone detection and summarization. Improving robustness of ML4Code.
Grounded Graph Decoding improves Compositional Generalization in Question Answering
EMNLP 2021 Empirical Methods in Natural Language Processing
We propose Grounded Graph Decoding, a method to improve compositional generalization of language representations by grounding structured predictions with an attention mechanism.
Accelerating Quadratic Optimization with Reinforcement Learning
NeurIPS 2021 Neural Information Processing Systems
We demonstrate reinforcement learning can significantly accelerate first-order optimization, outperforming state-of-the-art solvers by up to 3x.
Representing Long-Range Context for Graph Neural Networks with Global Attention
NeurIPS 2021 Neural Information Processing Systems
Transformers enable GNNs to achieve SOTA performance on graph classification tasks by enabling representations of long-range context.
Synthesizing Low-Power Approximate Hardware with Large-Scale Search
MLArchSys 2021 Workshop on Machine Learning in Architecture Systems as ISCA
Checkmate: Breaking the Memory Wall with Optimal Tensor Rematerialization
MLSys 2020 3rd Conference on Machine Learning and Systems
Use up to 5x less memory when training DNNs by recomputing activations.
The Case for GPU Multitenancy: The OoO VLIW JIT Compiler for GPU Inference
arXiv 2019 arXiv:1910.02653, Jan 2019
Demonstrate 2.5x-4.9x speedups for deep learning inference workloads via GPU multitenancy.
Revec: Program Rejuvenation through Revectorization
CC 2019 International Conference on Compiler Construction
Achieve performance portability for hand-vectorized programs, with up to a 1.88x speedup.
Dynamic Space-Time Scheduling for GPU Inference
LearningSys 2018 LearningSys workshop at NeurIPS 2018
Demonstrate 2.5x-4.9x speedups for deep learning inference workloads via GPU multitenancy.
DSCnet: Replicating Lidar Point Clouds with Deep Sensor Cloning
WAD 2018 Workshop on Autonomous Driving at CVPR
Generate high resolution LiDAR from ensembles of cheap sensors.
Scalable Architecture for Anomaly Detection and Visualization in Power Generating Assets
HPBDC 2017 HPBDC at IPDPS 2017
We present a scalable system for time-series anomaly detection that horizontally scales to sensor networks with >400k QPS.
Spotting Suspicious Reviews via (Quasi-)clique Extraction
S&P 2015 IEEE Security and Privacy (poster)
We uncover suspicious activity by well-organized reviewer rings sponsored by Yelp. Our work sheds light on Yelp's little-known paid review operations.