Generating Accurate Pseudo Examples for Continual Learning

Daniel L. Silver, Sazia MahfuzWorkshops @ CVPR 2020 · DOI

Abstract

Continual learning (CL) requires models to acquire new knowledge without forgetting previously learned tasks, a challenge known as catastrophic forgetting. While rehearsal-based methods mitigate this by re-training on old data, memory and privacy constraints often make storing original datasets unfeasible. This paper proposes a novel pseudo-rehearsal method using Restricted Boltzmann Machines (RBMs) to generate high-fidelity synthetic examples. By filtering uniform noise through a trained RBM and selecting samples with low reconstruction error, the method produces pseudo-examples that accurately adhere to the original data distribution. Experimental results on synthetic and real-world datasets demonstrate that this approach maintains model accuracy effectively, performing comparably to rehearsal with original data.

Introduction

The "stability-plasticity dilemma" is a fundamental hurdle in artificial intelligence. A model must be plastic enough to integrate new information but stable enough to preserve existing knowledge. In standard neural networks, new weight updates often overwrite previous ones, leading to catastrophic forgetting.

Pseudo-rehearsal offers a solution by using a generative model to produce "pseudo-examples" (PEs) representing forgotten tasks. However, the success of this approach hinges on the quality of the generated data. If PEs do not accurately reflect the underlying probability distribution of the original input space, the model's performance on prior tasks will degrade during consolidation.

Methodology

The proposed method leverages the generative capabilities of Restricted Boltzmann Machines (RBMs). Unlike standard sampling techniques, the authors introduce a filtering mechanism to ensure data quality.

RBM-Based Pseudo-Example Generation

  1. Noise Injection: Uniform random noise is passed through a trained RBM.
  2. Reconstruction Filtering: The RBM calculates the reconstruction error (Mean Squared Error or Euclidean distance) for each input.
  3. Selection: Only inputs with a reconstruction error lower than the mean training error of the original dataset are retained. This ensures the RBM "recognizes" the generated samples as belonging to the learned manifold.

Autoencoder Divergence Measure (ADM)

To evaluate the similarity between the original data and the generated PEs without requiring the storage of original datasets, the authors introduce the ADM. It is defined as the ratio of the reconstruction MSE of a test set (or PEs) to the reconstruction MSE of the original training set. An ADM value close to 1 indicates high distributional similarity.

Experimental Results

The method was validated across synthetic Gaussian clusters, the UCI Hypothyroidism dataset, and Boolean logic tasks.

Distributional Fidelity

In a 4D synthetic task involving multi-modal Gaussian distributions, the generated PEs achieved an ADM of 0.5772. Statistical validation using Maximum Mean Discrepancy (MMD) confirmed that the PEs were indistinguishable from the original training distribution.

Comparison of probability density functions (PDFs) for original training data (blue) and generated pseudo-examples (orange) across four dimensions.

As shown in the figure above, the RBM-filtered pseudo-examples (orange) closely track the complex, multi-modal probability density functions of the original training data (blue) across all four dimensions ($x_1$ through $x_4$).

Performance in Continual Learning

The effectiveness of these PEs was tested in a task consolidation scenario (Task A and Task B). A model using RBM-generated PEs for rehearsal (MTL:A+PE(B)) achieved 92.00% accuracy, nearly identical to the 91.83% accuracy of a model trained on the actual original data. In contrast, models rehearsed with unfiltered random noise saw a significant drop in performance to approximately 73%.

Conclusion

This research demonstrates that RBM-based filtering is a robust mechanism for generating accurate pseudo-examples for continual learning. By using reconstruction error as a proxy for probability, the method ensures that synthetic data remains faithful to the original task's distribution. Future work will explore the scalability of this approach to high-dimensional image data and its performance over longer sequences of tasks.

Summary updated

BibTeX
@inproceedings{silver2020generating,
  author = {Daniel L. Silver and Sazia Mahfuz},
  title = {{Generating Accurate Pseudo Examples for Continual Learning}},
  booktitle = {Workshops @ CVPR},
  year = {2020},
  doi = {10.1109/cvprw50498.2020.00136},
}
Figure