Source-linked AI summary
An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks
Ian J. Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, Yoshua Bengio
TL;DR
Catastrophic forgetting raises the question of how modern neural networks retain an earlier task after learning a new one. The paper compares training algorithms and activation functions across task relationships, finding dropout consistently best while activation rankings vary by task and tradeoff emphasis.
Problem
Catastrophic forgetting may cause learning systems to lose performance on an earlier task after training on a second task.
Method
The paper compares gradient-based training algorithms and activation functions across functionally identical, similar, and dissimilar task pairs.
Results
Dropout performs best for new-task performance, old-task performance, and the tradeoff between them, while activation-function rankings depend on task conditions.
Takeaways & Limitations
Activation functions should be cross-validated, while maxout with dropout is a practical choice when computational resources limit experimentation.
Abstract
from arXiv · showhide
Catastrophic forgetting is a problem faced by many machine learning models and algorithms. When trained on one task, then trained on a second task, many machine learning models "forget" how to perform the first task. This is widely believed to be a serious problem for neural networks. Here, we investigate the extent to which the catastrophic forgetting problem occurs for modern neural networks, comparing both established and recent gradient-based training algorithms and activation functions. We also examine the effect of the relationship between the first task and the second task on catastrophic forgetting. We find that it is always best to train using the dropout algorithm--the dropout algorithm is consistently best at adapting to the new task, remembering the old task, and has the best tradeoff curve between these two extremes. We find that different tasks and relationships between tasks result in very different rankings of activation function performance. This suggests the choice of activation function should always be cross-validated.
1. Introduction
Catastrophic forgetting can cause learning systems to lose performance on an earlier task after training on a second, motivating study of how task relationships and modern training choices affect retention. The paper finds dropout consistently performs best, while activation-function rankings depend on the task relationship and performance tradeoff.
- Catastrophic forgetting occurs when training on a second task causes a system to forget how to perform the first task.
- Complementary algorithms may be needed to reduce information loss from learning algorithms prone to forgetting.The paper motivates understanding forgetting before designing such complementary algorithms.
- The experiments compare functionally identical, similar, and dissimilar task relationships.The functionally identical tasks use different input formats.
- Dropout is consistently the best training algorithm for adapting to new tasks, retaining old-task performance, and balancing both objectives.
- Activation-function performance varies with task relationship and whether adaptation or retention receives greater emphasis.With dropout, maxout consistently appears somewhere on the performance-tradeoff frontier, but is not best at every point.
- Dropout often increases the optimal network size, although it usually decreases that size for dissimilar task pairs.The larger networks may explain some resistance to forgetting, while other beneficial dropout effects remain to be characterized.
2. Related work
Recent work had renewed interest in catastrophic forgetting in modern deep networks, but earlier experiments were limited in scope and methodology. This paper addresses those limitations with broader experiments across task relationships, hyperparameters, validation-based stopping, and training algorithms.
- Catastrophic forgetting received limited attention after the deep-learning renaissance, until later work repopularized its study.
- Earlier experiments trained one network per case with shared hyperparameters, heuristic stopping, one task pair, and standard gradient descent.
- The present study expands evaluation using multiple networks, varied hyperparameters, validation-based stopping, three task pairs, and dropout.
3. Methods
The experiments compare dropout and standard gradient-based training across neural-network activation functions, using varied task relationships and automated hyperparameter selection.
- Dropout: Dropout modifies stochastic gradient descent by randomly removing input and hidden units during training and scaling outgoing weights at test time.
- Dropout: Dropout can be viewed as training exponentially many weight-sharing neural networks and averaging their predictions, resembling bagging.
- Dropout: Dropout regularization allowed the best validation-selected MNIST network to use 56.48% more parameters than the best network without dropout.
- Activation functions: Each hidden layer computes z = Wx+b, applies an activation function h = f(z), and passes h to the next layer.
- Activation functions: The experiments use rectified linear units, hard local winner take all, and maxout among the studied activation functions.
- Methods: The study trains four activation functions with two algorithms, producing eight distinct methods for comparison.
- Random hyperparameter search: Random hyperparameter search is used instead of grid search to support fairer comparisons while limiting the number of experiments.
- Random hyperparameter search: The authors note that Bayesian optimization may perform better, but use random search because it achieved state-of-the-art performance on the considered tasks with 25 experiments.
4. Experiments
The experiments train networks sequentially on old and new tasks, comparing SGD and dropout across four activation functions and three task relationships. Across input-reformatted and semantically similar tasks, dropout improved two-task validation performance for all tested models, while the input-reformatted results suggest higher-layer adaptation.
- Each experiment trains a network on an old task, then trains it on a new task.
- The study compares SGD and dropout with logistic sigmoid, rectifier, hard LWTA, and maxout activations.
- Twenty-five randomly configured hyperparameter settings are evaluated for each of eight algorithm–activation conditions using two hidden layers and a softmax classifier.The search includes layer sizes, initialization, constraints, biases, and learning-rate and momentum schedules.
- Input reformatting: The input-reformatting tasks use MNIST with different pixel permutations, preserving digit concepts while changing each pixel’s meaning.The network must associate new pixel collections with penstrokes without disrupting higher-level concepts or old pixel–penstroke connections.
- Task relationships: Dropout improved two-task validation performance for all models on both the input-reformatted and similar task pairs.The similar-task experiment used sentiment classification across Amazon product categories, while another similar-task setup paired Amazon reviews with two-class MNIST.
- Input reformatting: For input reformatting, first-layer weight semantics changed little between tasks, suggesting higher layers adapted to the new input projection.
5. Discussion
Dropout consistently provides the strongest balance between adapting to new tasks and retaining performance on old tasks, while activation-function rankings depend on task conditions. The authors therefore recommend cross-validation when feasible and maxout with dropout when resources are limited.
- Dropout improves new-task performance, old-task performance, and tradeoff-curve points across all three task pairs.
- Dropout’s benefits may partly reflect larger trainable model sizes, although additional unexplained effects appear, especially across task relationships.
- Activation-function rankings are problem dependent, with logistic sigmoid ranging from worst under some conditions to best under others.
- Hard LWTA is not generally resistant to catastrophic forgetting and performs poorly in some task-pair and optimizer settings.
- Maxout trained with dropout is the only method appearing on the lower-left performance frontier for all three task pairs.