Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fortuitous Forgetting in Connectionist Networks

Hattie Zhou, Ankit Vani, Hugo Larochelle, Aaron Courville

TL;DR

The paper argues that forgetting can beneficially shape learning in neural networks, introducing the forget-and-relearn paradigm as a unifying lens for iterative training across image classification and language emergence. It formalizes a forgetting operation that lowers accuracy yet preserves information, and shows that relearning amplifies features that are robust across varying conditions. Through targeted forgetting methods like LLF and partial weight perturbations, the authors demonstrate improved generalization over baselines and provide mechanistic insights into why iterative retraining strengthens consistently useful representations. The work also extends the framework to emergent languages, analyzing compositionality and showing that balanced forgetting can enhance structured communication, with implications for designing more effective iterative training algorithms.

Abstract

Forgetting is often seen as an unwanted characteristic in both human and machine learning. However, we propose that forgetting can in fact be favorable to learning. We introduce "forget-and-relearn" as a powerful paradigm for shaping the learning trajectories of artificial neural networks. In this process, the forgetting step selectively removes undesirable information from the model, and the relearning step reinforces features that are consistently useful under different conditions. The forget-and-relearn framework unifies many existing iterative training algorithms in the image classification and language emergence literature, and allows us to understand the success of these algorithms in terms of the disproportionate forgetting of undesirable information. We leverage this understanding to improve upon existing algorithms by designing more targeted forgetting operations. Insights from our analysis provide a coherent view on the dynamics of iterative training in neural networks and offer a clear path towards performance improvements.

Fortuitous Forgetting in Connectionist Networks

TL;DR

The paper argues that forgetting can beneficially shape learning in neural networks, introducing the forget-and-relearn paradigm as a unifying lens for iterative training across image classification and language emergence. It formalizes a forgetting operation that lowers accuracy yet preserves information, and shows that relearning amplifies features that are robust across varying conditions. Through targeted forgetting methods like LLF and partial weight perturbations, the authors demonstrate improved generalization over baselines and provide mechanistic insights into why iterative retraining strengthens consistently useful representations. The work also extends the framework to emergent languages, analyzing compositionality and showing that balanced forgetting can enhance structured communication, with implications for designing more effective iterative training algorithms.

Abstract

Forgetting is often seen as an unwanted characteristic in both human and machine learning. However, we propose that forgetting can in fact be favorable to learning. We introduce "forget-and-relearn" as a powerful paradigm for shaping the learning trajectories of artificial neural networks. In this process, the forgetting step selectively removes undesirable information from the model, and the relearning step reinforces features that are consistently useful under different conditions. The forget-and-relearn framework unifies many existing iterative training algorithms in the image classification and language emergence literature, and allows us to understand the success of these algorithms in terms of the disproportionate forgetting of undesirable information. We leverage this understanding to improve upon existing algorithms by designing more targeted forgetting operations. Insights from our analysis provide a coherent view on the dynamics of iterative training in neural networks and offer a clear path towards performance improvements.
Paper Structure (29 sections, 4 equations, 12 figures, 11 tables)

This paper contains 29 sections, 4 equations, 12 figures, 11 tables.

Table of Contents

  1. Introduction
  2. The Forget-and-Relearn Hypothesis.
  3. Existing Algorithms as Instances of Forget-and-Relearn
  4. Image Classification.
  5. Language Emergence.
  6. Partial Weight Perturbation as Targeted Forgetting
  7. Image Classification
  8. Compositional language emergence
  9. Targeted Forgetting Improves Performance
  10. Reducing Overfitting in Image Classification
  11. Analysis of Related Work.
  12. Analysis of LLF.
  13. Increasing Compositionality in Emergent Languages
  14. Understanding the Forget-and-Relearn dynamic
  15. Image Classification
  16. ...and 14 more sections

Figures (12)

  • Figure 1: The left panels in \ref{['subfig:diff']} and \ref{['subfig:misl']} show training accuracy of the two example groups for two types of weight perturbation. The right panels in \ref{['subfig:diff']} and \ref{['subfig:misl']} show the accuracy of each example group for the same model with weight perturbation applied. Same colors represent the same forgetting operation, and dashed lines represent the accuracy on the difficult or mislabeled examples. Results are averaged over $5$ runs.
  • Figure 2: The left panels in \ref{['subfig:compperm_reinit']} and \ref{['subfig:compperm_zero']} show train accuracy of the compositional (Comp) and non-compositional (Perm) language examples; blue lines here are hidden behind orange lines. The right panels in \ref{['subfig:compperm_reinit']} and \ref{['subfig:compperm_zero']} show the accuracy of each group for the same model with weight perturbation applied. Same colors represent the same forgetting operation, and dashed lines represent the accuracy on the non-compositional examples. Results are averaged over $5$ runs.
  • Figure 3: Analysis experiments performed on Flower dataset with ResNet18. \ref{['fig:reverse']}-\ref{['fig:freeze']} show the min, max, and average of $3$ runs. \ref{['fig:zoomtrain']} shows that KE induces no forgetting after the first $2$ generations. \ref{['fig:reverse']} shows that resetting early layers significantly under-performs LLF. \ref{['fig:freeze']} shows that freezing early layers prevents any improvement in subsequent training, while freezing later layers still leads to significant improvements.
  • Figure 4: Topographic similarity ($\rho$) in the Lewis game with different forms of forgetting. \ref{['subfig:topofull']} presents the $\rho$ across all phases of iterated learning; \ref{['subfig:topozoomed']} zooms into the first 40000 steps to illustrate that $\rho$ improves in two stages for every generation: first during reinitialization and imitation (forget), but it is generally during interaction that the sender starts outperforming the previous generation's $\rho$ (relearn). \ref{['subfig:topoeot']} plots $\rho$ at the end of each generation for the ease-of-teaching setting.
  • Figure A1: The left panels in \ref{['conv4:easy']} and \ref{['resnet:easy']} show training accuracy of easy and hard example groups for different types of weight perturbation. The right panels in \ref{['conv4:easy']} and \ref{['resnet:easy']} show the accuracy of each example group for the same model with weight perturbation applied. Same colors represent the same forgetting operation, and dashed lines represent the accuracy on the hard examples. Results are averaged over $5$ runs. Lowess smoothing is applied for visual clarity.
  • ...and 7 more figures