PaperlayerView exact revision on arXiv

Source-linked AI summary

A Wholistic View of Continual Learning with Deep Neural Networks: Forgotten Lessons and the Bridge to Active and Open World Learning

Martin Mundt, Yongwon Hong, Iuliia Pliushch, Visvanathan Ramesh

arXiv:2009.01797v3cs.LGstat.ML

TL;DR

The paper addresses the closed-world evaluation of continual learning, where emphasis on catastrophic forgetting overlooks unknown and corrupted data. It critically surveys continual, active, and open set learning, proposes a consolidated view, and reports empirical support for improved forgetting mitigation, querying, task ordering, and open-world robustness. The authors also identify limitations in the specific neural-network realization and its EVT assumptions.

  • Problem

    Continual learning is commonly evaluated on closed-world benchmarks focused on catastrophic forgetting, despite neural networks producing false predictions on unknown and corrupted instances.

  • Method

    The paper critically surveys continual learning, active learning, and open set recognition, then develops a consolidated framework connecting their complementary perspectives.

  • Results

    The consolidated viewpoint is supported empirically across catastrophic-forgetting mitigation, active querying, task ordering, and robustness to open-world corruptions.

  • Takeaways & Limitations

    Continual learning evaluation should question whether current benchmarks realistically reflect systems that encounter unknown data and corruptions.

  • Takeaways & Limitations

    The specific neural-network method is not claimed to be generally best, and its EVT approach assumes uni-modal class clusters in latent space without a strict theoretical guarantee.

Abstract

from arXiv · show

Current deep learning methods are regarded as favorable if they empirically perform well on dedicated test sets. This mentality is seamlessly reflected in the resurfacing area of continual learning, where consecutively arriving data is investigated. The core challenge is framed as protecting previously acquired representations from being catastrophically forgotten. However, comparison of individual methods is nevertheless performed in isolation from the real world by monitoring accumulated benchmark test set performance. The closed world assumption remains predominant, i.e. models are evaluated on data that is guaranteed to originate from the same distribution as used for training. This poses a massive challenge as neural networks are well known to provide overconfident false predictions on unknown and corrupted instances. In this work we critically survey the literature and argue that notable lessons from open set recognition, identifying unknown examples outside of the observed set, and the adjacent field of active learning, querying data to maximize the expected performance gain, are frequently overlooked in the deep learning era. Hence, we propose a consolidated view to bridge continual learning, active learning and open set recognition in deep neural networks. Finally, the established synergies are supported empirically, showing joint improvement in alleviating catastrophic forgetting, querying data, selecting task orders, while exhibiting robust open world application.

1. Introduction

The paper argues that continual learning must move beyond preventing catastrophic forgetting on closed-world benchmarks. It proposes connecting continual learning with active learning and open set recognition to address unknown data and realistic deployment conditions.

  • Continual learning studies models that learn from data arriving over time rather than from a single previously collected dataset.
  • Current evaluations focus heavily on preserving representations and reducing catastrophic forgetting while often assuming deployment data matches the training distribution.
  • Unknown examples, dataset outliers, and common corruptions expose weaknesses in many deep continual learning systems that are evaluated under closed-world assumptions.
  • Real-world continual learning must distinguish known from novel instances and decide which newly arriving data is meaningful for future learning.
  • The paper develops a consolidated view that connects continual learning, active learning, and open set recognition, whose interactions are often overlooked when studied separately.

2. Preamble: continual machine learning

The preamble traces continual learning definitions and argues that its evaluation should account for task, data, and sequence choices rather than relying mainly on isolated-learning benchmarks. It also frames deep neural networks as the paper’s central setting.

  • Lifelong machine learning is presented as continuous learning across tasks that accumulates knowledge and uses it to support future learning.
  • Chen and Liu’s definition explicitly treats preserving accumulated information across tasks as a fundamental goal, unlike definitions focused only on helping the next task.
  • The literature lacks coherent descriptions of domain, formalizations of knowledge, and stable practices for choosing, ordering, and evaluating task sequences.
  • Deep neural networks make knowledge management difficult because learned representations are densely entangled within layers and distributed hierarchically across layers.
  • Continual learning benchmarks commonly split existing datasets into sequential task sets and compare performance with isolated training, while often disregarding task, permutation, and data choices.

3. Critically surveying and bridging three perspectives

The paper critically surveys continual learning, active learning, and open set recognition separately before identifying their shortcomings and presenting a consolidated view of their interconnections.

  • The survey categorizes methods and practices in continual learning, active learning, and open set recognition before examining their pitfalls and shortcomings.
  • Its consolidated view is based on the conjecture that open set recognition provides a natural interface between active and continual learning.

3.1. Continual learning

Continual learning is organized around preserving prior representations while incorporating arriving data. The section surveys regularization, rehearsal, and architectural approaches, including combinations of these strategies.

  • Continual learning methods are categorized into regularization, rehearsal, architectural, and combined approaches.The taxonomy follows established reviews while noting that strict separation is not always possible.
  • Regularization: Regularization balances protecting existing representations against retaining flexibility for new information.Regularization is especially intuitive when tasks share substantial structure in their high-dimensional embeddings.
  • Regularization: Knowledge distillation regularizes outputs toward stored soft targets, as in learning without forgetting and related approaches.The method was originally proposed for model compression and is often combined with other mechanisms.
  • Rehearsal: Rehearsal preserves information by replaying previously seen data, typically selecting a subset to limit memory and computation costs.Generative replay replaces stored examples with samples produced by a generative model.
  • Architectural approaches: Architectural approaches modify the network, using fixed-capacity routing or dynamic growth to reduce interference between representations.Dynamic-growth methods add representational units as new information is encountered.

3.2. Active learning

Active learning selects which unlabeled instances to label rather than focusing on preserving representations. The section reviews acquisition functions ranging from uncertainty heuristics to theoretically grounded and representation-based strategies.

  • Active learning queries a subset of unlabeled data expected to maximize performance gain after human annotation.The acquisition function determines which instances are selected from the unlabeled pool.
  • Most practical active-learning evaluations assume a closed world in which the complete unlabeled pool is accessible.Some approaches also incorporate the entire pool into training through semi-supervised learning.
  • Uncertainty-based approaches: Deep active-learning methods commonly use heuristics or approximations because uncertainty quantities are difficult to calculate in black-box neural networks.Examples include ensembles and Monte Carlo Dropout-based disagreement methods.
  • Version Space and Expected Error Reduction: Version-space methods select queries that reduce the set of hypotheses consistent with observed data, with guarantees established for some parametric models.Theoretical feasibility has been shown for Gaussian mixture models and naive Bayes, but not maturely established for deep neural networks.
  • Representation based approaches: Representation-based approaches select informative or diverse data using clustering, core sets, information density, or synthesized queries.Query-synthesizing methods rely on generative models capturing the data distribution sufficiently well to produce meaningful instances.

3.3. Open set recognition

Open set recognition addresses predictions on inputs outside the observed training distribution, where neural networks can be confidently wrong. The section distinguishes knowns from known and unknown unknowns and reviews methods for limiting such errors.

  • Open set recognition studies how models handle inputs beyond the closed world of the training distribution.It arose from observations that discriminative neural networks can make overconfident predictions on unseen data.
  • Input categories: Knowns come from the training distribution, known unknowns are recognized as uncertain, and unknown unknowns belong to unexplored distributions.The three categories describe increasing separation from the model’s observed concepts.
  • Formalization: The closed space encloses the training set, while the open space is the remainder of the input or feature space.This formalization frames unknown recognition as distinguishing observed regions from everything outside them.
  • Unknown unknowns: Adding anticipated background or unknown examples cannot generally cover unknown unknowns because deployment may contain unenumerated statistical deviations.Examples include illumination, acquisition, sensor, dirt, surface-marking, and noise differences.
  • Predictive Anomalies: Uncertainty estimates can fail to separate knowns and unknowns, with unseen data overlapping known data in confidence or entropy.The cited example used 50 Monte Carlo Dropout passes, yet unknown predictions remained substantially indistinguishable.
  • Meta-recognition: Meta-recognition uses a procedure over learned features to reduce confidence as examples move farther from the observed data population.The reviewed compact abating probability approach uses extreme value theory and a Weibull distribution.

3.4. Bridging perspectives: notable past insights and their synergies

The survey argues that continual learning should be understood alongside active learning and open set recognition rather than evaluated solely through closed-world forgetting benchmarks. It identifies limitations of uncertainty-based querying and benchmark design, then proposes open set recognition and EVT-based meta-recognition as a shared interface for robust learning.

  • Bridging perspectives: Continual learning, active learning, and open set recognition are usually studied separately, leaving their intersection and potential synergies largely unexplored.The authors propose treating them in conjunction under a broader view of continual learning.
  • Forgotten lessons from past literature: Closed-world benchmarks commonly evaluate forgetting on sequentialized known-data tasks while assuming task oracles and flawless active-query labels.Such protocols do not test robustness to unknown data, outliers, corruptions, or realistic labeling conditions.
  • Forgotten lessons from past literature: Uncertainty is not predictive of the open set because deep models can be overconfident and uncertainty estimates are uncalibrated outside the training distribution.Uncertainty-based selection can therefore query meaningless or uninformative outliers.
  • Forgotten lessons from past literature: Task order and data selection can substantially alter continual-learning results, yet benchmark studies often use arbitrary class-based sequences or insufficient randomization.The survey reports a major divergence when tasks are selected by similarity or dissimilarity to already observed data.
  • Bridging perspectives: Open set recognition can gauge the boundary between seen and unseen data, enabling continual-learning protection and active-learning queries that avoid uninformative noise.The paper presents open set recognition as a natural general interface between active and continual learning.
  • Bridging perspectives: EVT-based meta-recognition can use Weibull modeling and inverse sampling both to reject unknowns, balance active-learning exploration, and construct continual-learning rehearsal subsets.The method samples the heavy tail for statistically deviant but relevant active-learning data and samples the known distribution for rehearsal.

4. Highlighting natural synergies with empirical evidence

The paper supports its unified view with four quantitative experiment sets using a shared neural-network framework. These experiments examine continual exemplar selection, active queries, task ordering, and related open-world synergies.

  • Four quantitative experiments empirically investigate the paper’s proposed synergies across continual, active, and open-world learning.Each experiment is summarized in a dedicated subsection.
  • All experiments use the same unified neural-network framework, making shared mechanisms the key novelty rather than a particular realization.The authors note that this reuse is atypical in related works and previous reviews.
  • Continual exemplar selection compares mechanisms intended to prevent catastrophic forgetting.The proposed common EVT-based foundation is compared with conventionally employed techniques.
  • Active-query experiments investigate whether the proposed EVT-based foundation can support data-selection strategies.The passage states that the foundation surpasses techniques, but the supplied text truncates the comparison.
  • The practical framework combines a VAE realization with auxiliary EVT components for knowledge protection, querying, and unknown-data rejection or storage.The diagram distinguishes traditional VAE components from dashed EVT-based components and maps these functions by color.

4.1. One way to unite perspectives with deep

The proposed realization uses a classifying variational autoencoder and EVT to model the observed latent distribution and identify its closed space. The resulting probabilities support unknown-data handling, active querying, exemplar selection, and task-order assessment.

  • One way to unite perspectives with deep: The framework uses a generative model because discriminative activation patterns and closed-space boundaries are not necessarily sufficient to identify unseen data.The passage motivates a VAE-based realization for modeling the data distribution more explicitly.
  • One way to unite perspectives with deep: A probabilistic encoder learns q(z|x), while a probabilistic decoder models p(x,z)=p(x|z)p(z) and supplies the generative component.The generative model can also be augmented to discriminate classes.
  • One way to unite perspectives with deep: Training jointly minimizes divergence from a chosen prior together with reconstruction and linear classification objectives.The supplied passage describes these objectives as parametrized through θ, ϕ, and ξ.
  • 4.1.1. The boundary between known and unknown: The aggregate posterior is treated as the empirically captured data distribution rather than being assumed identical to the prior.The β factor controls the prior–approximate-posterior gap, while the encoding is intended to preserve both class separation and correlations.
  • 4.1.1. The boundary between known and unknown: EVT fits class-conditional latent-distance distributions to approximate trustworthy high-density regions and identify the model’s closed space.The distances may use Euclidean or cosine metrics, and Weibull parameters model the corresponding regions.
  • 4.1.2. Approximate posterior based open set recognition: CDF-based outlier probabilities enable unknown-data identification, rejection, or storage, while the minimum across known classes determines the closest known-data deviation.A sample is anomalous only when its outlier probability is large for every known class.
  • 4.1.3. Active data selection: For active learning, small CDF values indicate similarity to existing representations, whereas larger values indicate previously unobserved data.The proposed query strategy samples across the center of the CDF rather than selecting only the most outlying points, balancing informativeness with robustness to noisy or uninformative data.
  • 4.1.4. Continual exemplar selection: Continual learning uses the same representation to select compact exemplars while excluding statistical anomalies, supporting robust deployment after learning.Exemplar quality depends on the desired quantity, with outlier probabilities limited, for example, to p<0.95.

4.2. Empirical evidence

The experiments examine core-set selection, active querying, corruption robustness, task ordering, and evaluation choices in continual learning. Across these settings, EVT-based selection and broader experimental controls improve or materially affect observed performance.

  • Exemplar selection and core set extraction: Mini-batch balancing produced continual-learning accuracy gaps exceeding 5%, complicating comparisons between core-set methods with similar sizes.The study contrasts unbalanced, exemplar-balanced, and class-balanced sampling.
  • Exemplar selection and core set extraction: The latent EVT approach outperformed competing core-set selection methods as core-set sizes increased, while latent herding increasingly selected redundant samples.The authors attribute herding’s widening gap to its reliance on mean approximation without an accompanying diversity objective.
  • Exemplar selection and core set extraction: EVT-based core-set construction more closely approximated the aggregate posterior than greedy k-center selection in a two-dimensional CIFAR10 visualization.The improved distributional fit is presented as instrumental to long-term learning through core-set selection, not as the primary objective.
  • Active queries: In active learning, the latent EVT method significantly outperformed all evaluated methods throughout MNIST and CIFAR10 experiments.Its strategy balances novel and still-informative examples while rejecting uninformative outliers.
  • Robustness to open world corruptions: Under corrupted data, random sampling and EVT-based querying remained robust, whereas uncertainty, reconstruction-loss, and k-means strategies were vulnerable to corrupted outliers.The continual-learning comparison also found robustness for random sampling, latent herding, and EVT-based selection, with EVT outperforming the other approaches.
  • The effect of task order and curricula: Task order changed continual-learning accuracy substantially: dissimilar-task selection improved performance, while proximity-based ordering reduced it on CIFAR100 and AudioMNIST.The authors hypothesize that nearby concepts create immediate classifier or generative-model confusion.
  • The effect of task order and curricula: Different CIFAR100 task orders produced absolute performance discrepancies as large as 10%, highlighting sensitivity relevant to reproducibility and benchmark significance.The authors argue that continual-learning evaluation should move beyond rigid benchmarks toward more holistic approaches.

5. Discussion: towards wholistic deep continual learning

The discussion broadens continual learning into a framework that incorporates open-world robustness, active data selection, and task curricula while acknowledging limitations of the proposed VAE implementation and current benchmarks.

  • Unified viewpoint: The proposed framework treats continual learning, active learning, and open set recognition as complementary components of a unified viewpoint.The authors argue that these paradigms have usually been studied separately despite potential synergies.
  • Revised definition: The authors propose extending continual learning definitions to include unknown-data recognition, active queries, representation transfer, and curriculum construction.The revised definition retains abstraction while adding factors considered relevant to realistic continual learning.
  • Prospects: The discussion identifies newer continual-learning datasets and more exhaustive evaluation protocols as necessary for investigating realistic applications.The authors acknowledge that their own experiments deliberately use familiar pseudo-continual datasets such as CIFAR10/100.
  • Limitations: The particular neural-network method is not claimed to be generally optimal, because alternative practical frameworks with the same scope remain insufficiently explored.The paper’s main goal is to highlight the consolidated viewpoint rather than establish universal superiority of its implementation.
  • Limitations: The VAE-based framework assumes unimodal per-class latent clusters for EVT, without a strict theoretical guarantee, especially in higher-dimensional settings.Multiple sub-clusters could occur on one side of a decision boundary, although this was not observed in the reported teaching examples.
  • Limitations: Conventional VAEs may be outperformed on large datasets, so scaling the framework to larger real-world data requires additional advances that can increase computational expense.The cited improvements include autoregression, adversarial training, and introspection.

6. Conclusion

The conclusion presents a common viewpoint uniting robust continual and active learning in the presence of unknown data. Experiments provide empirical support for benefits across the paper’s investigated aspects and motivate future adoption of the framework.

  • Conclusion: The paper unites robust continual learning and active learning while explicitly addressing unknown data.The authors describe this as a common viewpoint and report experimental investigations for each aspect.
  • Conclusion: The authors encourage future work to adopt the framework and investigate additional facets and applications.They note that the presented experiments can be extended with multiple facets and nuanced applications.
Loading 2009.01797v3…