ADA Lab @ UCSD

Project Krypton

Overview

Deep CNNs are now the preferred way to perform image analytics in many domains including healthcare, e-commerce, security, and sociology. However, one of the main criticisms pointed against deep CNNs is the black box nature of how they make predictions. To explain CNN predictions one of the widely used approaches in the practical literature is the occlusion based explanation approach (OBE for short). However, OBE experiments are highly time consuming as they need to perform large number of re-inference requests.

In this project we apply incremental and approximate inference optimizations to accelerate the OBE workload. Our work is inspired by the long line of work in incremental view maintenance, multi-query optimization, and approximate query processing techniques in the context of relational data management systems.

Downloads (Paper, Code, Data, etc.)

• Query Optimization for Faster Deep CNN Explanations
  Supun Nakandala, Arun Kumar, and Yannis Papakonstantinou
  ACM SIGMOD Record 2020 | Paper PDF and BibTeX
  ACM SIGMOD Research Highlights Award

• Incremental and Approximate Computations for Accelerating Deep CNN Inference
  Supun Nakandala, Kabir Nagrecha, Arun Kumar, and Yannis Papakonstantinou
  ACM TODS 2020 | Paper PDF and BibTeX
  Invited Paper

• Incremental and Approximate Inference for Faster Occlusion-based Deep CNN Explanations
  Supun Nakandala, Arun Kumar, and Yannis Papakonstantinou
  ACM SIGMOD 2019 | Paper PDF and BibTeX | TechReport | Blog post | Talk Video
  Honorable Mention for Best Paper Award

• Demonstration of Krypton: Optimized CNN Inference for Occlusion-based Deep CNN Explanations
  Allen Ordookhanians, Xin Li, Supun Nakandala, and Arun Kumar
  VLDB 2019 | Paper PDF and BibTeX | Video
  

• Demonstration of Krypton: Incremental and Approximate Inference for Faster Occlusion-based Deep CNN Explanations
  Supun Nakandala, Arun Kumar, and Yannis Papakonstantinou
  SysML 2019 Demo | Paper PDF | Video

Student Contact

Supun Nakandala: snakanda [at] eng [dot] ucsd [dot] edu

Acknowledgments

This project was/is supported in part by a Hellman Fellowship and by the NIDDK of the NIH under award number R01DK114945.