Timezone: »

Can Neural Network Memorization Be Localized?
Pratyush Maini · Michael Mozer · Hanie Sedghi · Zachary Lipton · Zico Kolter · Chiyuan Zhang

Thu Jul 27 01:30 PM -- 03:00 PM (PDT) @ Exhibit Hall 1 #323
Event URL: https://pratyushmaini.github.io/mem_web/ »

Recent efforts at explaining the interplay of memorization and generalization in deep overparametrized networks have posited that neural networks memorize ``hard'' examples in the final few layers of the model. Memorization refers to the ability to correctly predict on atypical examples of the training set. In this work, we show that rather than being confined to individual layers, memorization is a phenomenon confined to a small set of neurons in various layers of the model. First, via three experimental sources of converging evidence, we find that most layers are redundant for the memorization of examples and the layers that contribute to example memorization are, in general, not the final layers. The three sources are gradient accounting (measuring the contribution to the gradient norms from memorized and clean examples), layer rewinding (replacing specific model weights of a converged model with previous training checkpoints), and retraining (training rewound layers only on clean examples). Second, we ask a more generic question: can memorization be localized anywhere in a model? We discover that memorization is often confined to a small number of neurons or channels (around 5) of the model. Based on these insights we propose a new form of dropout---example-tied dropout that enables us to direct the memorization of examples to an aprior determined set of neurons. By dropping out these neurons, we are able to reduce the accuracy on memorized examples from 100% to 3%, while also reducing the generalization gap.

Author Information

Pratyush Maini (Carnegie Mellon University)
Michael Mozer (Google Research)
Hanie Sedghi (Google Deepmind)
Hanie Sedghi

Hanie Sedghi a Senior Research Scientist at Google DeepMind where she leads the DeepPhenomena team. The focus of her research has been understanding deep learning models to push their boundaries; not just for (out-of-distribution) generalization, but also the broader sense of algorithmic and scientific reasoning capabilities (of large language models). She is a workshop chair for NeurIPS 2022 as well as tutorial chair for ICML 2022 and 2023, a program chair for CoLLAs 2023 and has been an area chair for NeurIPS, ICLR and ICML and a member of JMLR Editorial board for the last few years. Prior to Google, Hanie was a Research Scientist at Allen Institute for Artificial Intelligence and before that, a postdoctoral fellow at UC Irvine. She received her PhD from University of Southern California with a minor in Mathematics.

Zachary Lipton (CMU & Abridge)
Zico Kolter (Carnegie Mellon University / Bosch Center for AI)
Chiyuan Zhang (MIT)

More from the Same Authors