`

Timezone: »

 
Poster
Multi-Stage Influence Function
Hongge Chen · Si Si · Yang Li · Ciprian Chelba · Sanjiv Kumar · Duane Boning · Cho-Jui Hsieh

Tue Dec 08 09:00 PM -- 11:00 PM (PST) @ Poster Session 2 #639

Multi-stage training and knowledge transfer, from a large-scale pretraining task to various finetuning tasks, have revolutionized natural language processing and computer vision resulting in state-of-the-art performance improvements. In this paper, we develop a multi-stage influence function score to track predictions from a finetuned model all the way back to the pretraining data. With this score, we can identify the pretraining examples in the pretraining task that contribute most to a prediction in the finetuning task. The proposed multi-stage influence function generalizes the original influence function for a single model in (Koh &Liang, 2017), thereby enabling influence computation through both pretrained and finetuned models. We study two different scenarios with the pretrained embedding fixed or updated in the finetuning tasks. We test our proposed method in various experiments to show its effectiveness and potential applications.

Author Information

Hongge Chen (MIT)
Si Si (Google Research)
Yang Li (Google)

Yang Li is a Senior Staff Research Scientist at Google, and an affiliate faculty member at the University of Washington CSE, focusing on the area intersecting AI and HCI. He pioneered on-device interactive ML on Android by developing impactful product features such as next app prediction and Gesture Search. Yang has extensively published in top venues across both the HCI and ML fields, including CHI, UIST, ICML, ACL, EMNLP, CVPR, NeurIPS (NIPS), ICLR, and KDD, and has constantly served as area chairs or senior area (track) chairs across the fields. Yang is also an editor of the upcoming Springer book on "AI for HCI: A Modern Approach", which is the first thorough treatment of the topic.

Ciprian Chelba (Google)
Sanjiv Kumar (Google Research)
Duane Boning (Massachusetts Institute of Technology)
Cho-Jui Hsieh (UCLA)

More from the Same Authors