You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
<h3class="tw-text-xl tw-font-medium tw-mt-4">Replay Can Provably Increase Forgetting</h3>
160
+
<pclass="tw-text-gray-600 tw-mt-4">
161
+
We provide a theoretical analysis of sample replay in over-parameterized continual linear regression, and we show that replay can provably increase forgetting in the worst case even though the network has the capacity to memorize all tasks.
<h3class="tw-text-xl tw-font-medium tw-mt-4">Replay Can Provably Increase Forgetting</h3>
248
+
<pclass="tw-text-gray-600 tw-mt-4">
249
+
We provide a theoretical analysis of sample replay in over-parameterized continual linear regression, and we show that replay can provably increase forgetting in the worst case even though the network has the capacity to memorize all tasks.
<h3class="tw-text-xl tw-font-medium tw-mt-4">Self-Supervised Learning of Video Representations from a Child's Perspective</h3>
486
-
<pclass="tw-text-gray-600 tw-mt-4">
487
-
We train self-supervised video models on longitudinal, egocentric headcam recordings collected from a child over a two year period in their early development.
<h3class="tw-text-xl tw-font-medium tw-mt-4">Learning and Forgetting Unsafe Examples in Large Language Models</h3>
520
-
<pclass="tw-text-gray-600 tw-mt-4">
521
-
We explore the behavior of LLMs finetuned on noisy custom data containing unsafe content and propose a simple filtering algorithm for detecting harmful content based on the phenomenon of selective forgetting.
Copy file name to clipboardExpand all lines: people/alex-wang/index.html
+2-2Lines changed: 2 additions & 2 deletions
Original file line number
Diff line number
Diff line change
@@ -2,7 +2,7 @@
2
2
<htmllang="en">
3
3
<head>
4
4
<metaproperty="og:title" content="Alex N. Wang | Agentic Learning AI Lab" />
5
-
<metaproperty="og:description" content="Alex Wang is a PhD student in CILVR advised by Mengye Ren at New York University. He did his MSc with Rich Zemel and BASc in Engineering Science at University of Toronto. He is supported by the NSERC PGS-D scholarship. He is interested in intelligent, large-scale generative vision models. Through generative modeling, he aims to define, build, and evaluate a notion of visual intelligence." />
5
+
<metaproperty="og:description" content="Alex Wang is a PhD student in CILVR at New York University. He did his MSc with Rich Zemel and BASc in Engineering Science at University of Toronto. He is supported by the NSERC PGS-D scholarship. He is interested in intelligent, large-scale generative vision models. Through generative modeling, he aims to define, build, and evaluate a notion of visual intelligence." />
Alex Wang is a PhD student in CILVR advised by Mengye Ren at New York University. He did his MSc with Rich Zemel and BASc in Engineering Science at University of Toronto. He is supported by the NSERC PGS-D scholarship. He is interested in intelligent, large-scale generative vision models. Through generative modeling, he aims to define, build, and evaluate a notion of visual intelligence.
118
+
Alex Wang is a PhD student in CILVR at New York University. He did his MSc with Rich Zemel and BASc in Engineering Science at University of Toronto. He is supported by the NSERC PGS-D scholarship. He is interested in intelligent, large-scale generative vision models. Through generative modeling, he aims to define, build, and evaluate a notion of visual intelligence.
<h3class="tw-text-xl tw-font-medium tw-mt-4">Replay Can Provably Increase Forgetting</h3>
126
+
<pclass="tw-mt-4">
127
+
Authors: Yasaman Mahdaviyeh, James Lucas, Mengye Ren, Andreas S. Tolias, Richard Zemel, and Toniann Pitassi
128
+
</p>
129
+
<pclass="tw-text-gray-600 tw-mt-4">
130
+
Continual learning seeks to enable machine learning systems to solve an increasing corpus of tasks sequentially. A critical challenge for continual learning is forgetting, where the performance on previously learned tasks decreases as new tasks are introduced. One of the commonly used techniques to mitigate forgetting, sample replay, has been shown empirically to reduce forgetting by retaining some examples from old tasks and including them in new training episodes. In this work, we provide a theoretical analysis of sample replay in an over-parameterized continual linear regression setting, where each task is given by a linear subspace and with enough replay samples, one would be able to eliminate forgetting. Our analysis focuses on sample replay and highlights the role of the replayed samples and the relationship between task subspaces. Surprisingly, we find that, even in a noiseless setting, forgetting can be non-monotonic with respect to the number of replay samples. We present tasks where replay can be harmful with respect to worst-case settings, and also in distributional settings where replay of randomly selected samples increases forgetting in expectation. We also give empirical evidence that harmful replay is not limited to training with linear models by showing similar behavior for a neural networks equipped with SGD. Through experiments on a commonly used benchmark, we provide additional evidence that, even in seemingly benign scenarios, performance of the replay heavily depends on the choice of replay samples and the relationship between tasks.
0 commit comments