Timezone: »
With the rise of pretrained models, fine-tuning has become increasingly important. However, naive fine-tuning often does not eliminate a model's sensitivity to spurious cues. To understand and address this limitation, we study the geometry of neural network loss landscapes through the lens of mode-connectivity. We tackle two questions: 1) Are models trained on different distributions mode-connected? 2) Can we fine tune a pre-trained model to switch modes? We define a notion of mechanistic similarity based on shared invariances and show linearly-connected modes are mechanistically similar. We find naive fine-tuning yields linearly connected solutions and hence is unable to induce relevant invariances. We also propose and validate a method of "mechanistic fine-tuning" based on our gained insights.
Author Information
Ekdeep S Lubana (University of Michigan; CBS, Harvard University)
Eric Bigelow (Harvard University)
Robert Dick (University of Michigan)
David Krueger (University of Cambridge)
Hidenori Tanaka (Harvard University, Harvard University)
More from the Same Authors
-
2021 : Multi-Domain Balanced Sampling Improves Out-of-Distribution Generalization of Chest X-ray Pathology Prediction Models »
Enoch Tetteh · David Krueger · Joseph Paul Cohen · Yoshua Bengio -
2022 : Domain Generalization for Robust Model-Based Offline Reinforcement Learning »
Alan Clark · Shoaib Siddiqui · Robert Kirk · Usman Anwar · Stephen Chung · David Krueger -
2022 : Domain Generalization for Robust Model-Based Offline RL »
Alan Clark · Shoaib Siddiqui · Robert Kirk · Usman Anwar · Stephen Chung · David Krueger -
2022 : What shapes the loss landscape of self-supervised learning? »
Liu Ziyin · Ekdeep S Lubana · Masahito Ueda · Hidenori Tanaka -
2022 : Geometric Considerations for Normalization Layers in Equivariant Neural Networks »
Max Aalto · Ekdeep S Lubana · Hidenori Tanaka -
2022 : On The Fragility of Learned Reward Functions »
Lev McKinney · Yawen Duan · Adam Gleave · David Krueger -
2022 : Training Equilibria in Reinforcement Learning »
Lauro Langosco · David Krueger · Adam Gleave -
2022 : Assistance with large language models »
Dmitrii Krasheninnikov · Egor Krasheninnikov · David Krueger -
2022 : A Mechanistic Lens on Mode Connectivity »
Ekdeep S Lubana · Eric Bigelow · Robert Dick · David Krueger · Hidenori Tanaka -
2022 : Assistance with large language models »
Dmitrii Krasheninnikov · Egor Krasheninnikov · David Krueger -
2022 : Unifying Grokking and Double Descent »
Xander Davies · Lauro Langosco · David Krueger -
2022 Poster: Defining and Characterizing Reward Gaming »
Joar Skalse · Nikolaus Howe · Dmitrii Krasheninnikov · David Krueger -
2022 Poster: Analyzing Data-Centric Properties for Graph Contrastive Learning »
Puja Trivedi · Ekdeep S Lubana · Mark Heimann · Danai Koutra · Jayaraman Thiagarajan -
2021 Poster: Beyond BatchNorm: Towards a Unified Understanding of Normalization in Deep Learning »
Ekdeep S Lubana · Robert Dick · Hidenori Tanaka