acceptodds
Withdrawn from ICLR 2027

Evidence-Driven Progressive Knowledge Formation and Selective Adaptation: A Neuroscience-Inspired Continual Learning Framework

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Current object detectors suffer from catastrophic forgetting in open-world continual learning, where novel objects emerge progressively from incomplete, noisy, or ambiguous observations. Existing methods typically adapt representations directly to incoming observations, causing immature novel concepts to be prematurely assimilated into established knowledge and destabilizing previously learned representations. Neuroscience suggests that lifelong learning relies on complementary memory systems, where specific experiences are progressively consolidated into generalized knowledge while established representations remain relatively stable. Inspired by this principle, we propose Evidence-driven Progressive Knowledge formation and selective Adaptation (EPKA), a new continual learning framework that transforms uncertain observations into stable knowledge before allowing them to reshape existing representations. EPKA comprises two coupled mechanisms. Progressive Structural Evidence Evolution (PSEE) progressively accumulates and consolidates compatible observations into persistent, structurally coherent novel representations. Adaptive Representation Competition Regulation (ARCR) selectively adapts representations based on their stability, while protecting established knowledge from unnecessary changes. Together, they establish an evidence-knowledge-adaptation process that balances plasticity for emerging concepts with stability of established knowledge. Extensive experiments across three open-world continual learning scenarios and five benchmarks demonstrate that EPKA consistently improves novel knowledge acquisition while mitigating catastrophic forgetting, highlighting its adaptability and plasticity in continuously evolving visual environments.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Continual Learning, Progressive Knowledge Formulation, Selective Adaption, Evidence-driven, Plasticity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9oqUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper claims a novel continual learning framework for open-world object detection by decoupling preogressive knowledge formation from selective represenation adaptation. Initial experiments on multiple benchmarks supports the efficacy of the proposed framework.

Critical strengths

None

Critical weaknesses

The serious issue lies in the presentation of the methodology. In particular, while various modules and function are introduced in the framework, there are no any description and justificaiton for them. For example, how to project the feature into spatial and frequency response spaces? how can spatial and frequency features be directly sumed considering that they are from two different two different spaces? what is adaptive Fourier-domain refinement? how to get E^s and E^f in Eq (1)? why keep smoothness and sparsity for E (Eq 2)? what are teh symboles appearing in Eq (3)? what is LIF-style temporal state? What is relationship between Eq (3) and E*{i,t} in Eq (2)? what is the temporal state v_t in Eq (4)? how to integrate to get S{i,t}? what is phi_N in Eq (5)? what are "latent representation pathways"? what inputs are associated with "consecutive predictions"? what is g_N in Eq (8)? what is h_{i,t} for in Eq (9)? what are the structure of the mentioned mappings/functions?

Considering the very ambgious description of the methodology, the validity of experimental results are hard to evaluate. In addition, the sensitivity study of various hyper-parameteres are not valid, considering that only about three different values were set to each hyper-parameters with clear performance variations.

Questions for authors

Please well describe and justify the role, meaning, and structure of each module and function appearing the proposed framework.

Reviewer_gnDCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EPKA, a neuroscience-inspired continual learning framework for open-world object detection that separates progressive knowledge formation from subsequent representation adaptation. EPKA consists of two main components: Progressive Structural Evidence Evolution (PSEE), which constructs, refines, and temporally consolidates evidence from emerging observations, and Adaptive Representation Competition Regulation (ARCR), which regulates representation adaptation according to estimated novelty and historical prediction stability. The framework is evaluated on five object-detection benchmarks spanning logo, fine-grained pest, and remote-sensing scenarios. The experiments report improvements in both unknown-object discovery and known-category recognition. Overall, the empirical results demonstrate the effectiveness of EPKA.

Critical strengths
  1. The paper studies an interesting and relevant problem in open-world continual learning by shifting attention from only preserving established knowledge to the reliability of emerging knowledge before adaptation. The distinction between progressive knowledge formation and selective representation adaptation is conceptually interesting, and PSEE and ARCR form a coherent overall framework.

  2. The empirical evaluation is reasonably broad, covering five benchmarks across three different visual domains, and EPKA achieves strong improvements in unknown-object discovery and known-category detection, particularly on the logo benchmarks.

Critical weaknesses
  1. The central hypothesis that premature assimilation of unreliable observations is an important source of continual-learning instability is not directly evaluated under controlled variations of observation reliability, such as evidence sparsity, noise, occlusion, or class ambiguity. Moreover, key concepts such as reliable evidence and stable knowledge are not quantitatively defined or measured, making it difficult to determine whether the proposed evidence–knowledge–adaptation process actually behaves as claimed.

  2. The reliability interpretation of the constrained evidence refinement objective is insufficiently justified. The discussion following Eq. (2) assigns specific functional roles to the fidelity, smoothness, sparsity, and amplitude constraints and further argues that their combination produces structured evidence that is sufficiently reliable to persist. These explanations are overly intuitive and lack theoretical support derived from the optimization objective.

  3. The dynamic analysis provides limited empirical validation of the proposed mechanisms. Figure 3 presents a qualitative visualization of a representative unseen category and suggests that PSEE produces increasingly compact representations, but this behavior is not quantified across classes or datasets. Figure 4 appears primarily schematic rather than a quantitative experimental analysis: it illustrates the intended behavior of ARCR, but does not report measured pathway-wise update magnitudes, or plasticity-stability scores. Therefore, the current figures are useful for illustrating the proposed intuition, but provide limited evidence that PSEE and ARCR actually operate through the claimed mechanisms.

Questions for authors
  1. Could the authors directly validate the central premature-assimilation hypothesis by systematically varying observation reliability, such as evidence sparsity, corruption, or class ambiguity? In particular, does the advantage of EPKA increase as early evidence becomes less reliable?

  2. How are evidence reliability and stable knowledge formally defined in the paper? Could the authors quantitatively measure how these quantities evolve as observations accumulate, for example through representation drift, intra-class compactness, or inter-class separation?

  3. The appendix analyzes only a subset of the method's hyperparameters. Could the authors provide sensitivity analyses for the key parameters controlling temporal consolidation and ARCR, particularly the novelty–stability trade-off and warm-up behavior, and clarify how these parameters were selected?

  4. Could the authors report more direct continual learning measures such as per-task forgetting or BWT? This would help establish whether the improvements in previously known categories correspond to statistically robust reductions in catastrophic forgetting.

Reviewer_zF4KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a continual learning algorithm that is inspired by neuroscience. The method selectively adapts the model to integrate knowledge while preserving learned representations, which prevents models from learning noisy signals. Through experimental validation on open-world object detection and ablation studies, the paper demonstrates the proposed method's effectiveness.

Critical strengths
  1. The proposed method's motivation and connection to neuroscience are strong.
  2. The paper includes ablation studies of the proposed method, which clearly reflect the effectiveness of each component.
Critical weaknesses
  1. My biggest concern is the scope of this paper. While the method's high-level structure (Figure 1) is not specific to any domain, all experiments are conducted on open-world object detection. It is unclear why the paper focuses only on this scenario, and whether the proposed method is effective in other scenarios, such as language modeling and classification tasks.

  2. The baseline comparison, benchmarks, and evaluation metrics are rarely seen in the continual learning literature. Without common baselines and evaluation, it is unclear whether the contribution matters to the continual learning community.

  • The main results are not compared with common continual learning methods that are also related to neuroscience [1] or using memory [2]. Meanwhile, several methods in Table 1 are not even designed for continual learning, such as PROB and RandomBox.
  • Meanwhile, several baselines in Table 1 are not cited in the paper. It is hard for readers to check the papers.
  • The main results show the per-task performance for all 4 tasks, but do not measure standard metrics [2] like Average Accuracy, Backward Transfer, or Forward Transfer.

While the proposed method is well-motivated and the presentation is clear, I think the impact of this paper is limited because of these two reasons.


Reference:

[1] Overcoming catastrophic forgetting in neural networks, PNAS 2017

[2] Gradient Episodic Memory for Continual Learning, NIPS 2017

Questions for authors

The main questions are included in the Critical Weakness part.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.