PaperlayerView exact revision on arXiv

Source-linked AI summary

Brain-Inspired Hierarchical Modularity for General Continual Learning

Hongwei Yan, Kanglei Zhou, Qi Cheng, Weiyi Dong, Chunyan Lan, Guanglong Sun, Jun Zhou, Qian Li, Yi Zhong, Liyuan Wang

arXiv:2609.25146v1cs.LGcs.AI

TL;DR

FlyGCL addresses continual learning from online, uncertain, evolving streams, where systems must separate conflicting experience and integrate compatible experience. It coordinates expert specialization with ensemble integration in pretrained foundation models, improving performance across diverse settings, with especially large gains in embodied manipulation.

  • Problem

    General continual learning must learn from online, uncertain, evolving streams while separating conflicting experience and integrating compatible experience.

    Source: §1 · p. 2, §4.1 · p. 17

  • Method

    FlyGCL adapts pretrained foundation models with hierarchical modularity, combining random-expanded expert routing with differentiated modules and multi-timescale ensemble integration.

    Source: §1 · p. 3, §4.3 · p. 20

  • Results

    FlyGCL consistently improves performance across visual recognition, vision-language, ego-exo video, and embodied learning; in embodied manipulation it exceeds the strongest replay-free baseline by 49.5–58.8 percentage points across benchmarks.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The findings support hierarchical modularity as a unified principle for learning from dynamic experience by coordinating separation of conflicting experience with integration of compatible experience.

    Source: §3 · pp. 14–15

    FlyGCL concentrates plasticity in lightweight modules over relatively stable pretrained representations, leaving deeper learning within large foundation models beyond scope.

    Source: §3 · p. 16

Abstract

from arXiv · show

Continual learning, the ability to learn from sequential experience while retaining and adapting prior knowledge, is central to intelligent systems operating in changing environments. However, conventional continual learning is typically studied with offline task-wise training and clear task boundaries, leaving a substantial gap from general continual learning under online, uncertain, and evolving data streams. In this regime, intelligent systems must separate conflicting experience to reduce interference while integrating compatible experience to promote generalization. Inspired by the organization of the Drosophila learning and memory system, we identify a hierarchical modular principle that coordinates both functions through expert specialization and ensemble integration. We instantiate this principle as lightweight modular adaptation of pretrained foundation models, combining brain-inspired random expansion for expert routing and diversified modular integration across spatial and temporal scales. Across visual recognition, vision-language understanding, ego-exo video understanding, and embodied vision-language-action learning, our method consistently improves learning under online and uncertain data streams, with gains exceeding 50 percentage points over replay-free alternatives in embodied manipulation. These findings support hierarchical modularity as a biologically grounded path for learning from dynamic experience.

1 Introduction

General continual learning (GCL) concerns online, uncertain, evolving experience, where learning must separate conflicting data and integrate compatible data. FlyGCL draws on Drosophila-inspired hierarchical modularity and improves performance across continual-learning scenarios, especially embodied manipulation.

1 Introduction

  • Continual learning is framed as important for AI systems that learn from experience while adapting to changing conditions despite forgetting and loss of plasticity.The introduction connects this motivation to perspectives on an “era of experience” and self-improving agents.Source: §1 · p. 2
  • GCL extends conventional offline, task-wise learning with clear boundaries and auxiliary task identities to online, uncertain, overlapping, and evolving experience.Its central challenge is separating conflicting experience to reduce interference while integrating compatible experience to promote generalization.Source: §1 · p. 2
  • Drosophila olfactory learning combines sparse, largely random sensory projections with downstream memory pathways differentiated across spatial and temporal characteristics.The paper relates these functions to expert specialization and ensemble integration, coordinated hierarchically rather than used independently.Source: §1 · pp. 2–3
  • FlyGCL keeps a pretrained backbone as a stable representational substrate and uses random expansion for expert routing, with differentiated modules and multi-timescale predictions for integration.The framework is designed for lightweight, parameter-efficient learning across pretrained backbones and learning settings.Source: §1 · p. 3
  • FlyGCL achieves final average success rates of 83.1%, 86.1%, 94.5%, and 79.1% across four embodied manipulation settings, exceeding the strongest replay-free CL baseline by 49.5–58.8 percentage points.The settings are spatial, object-centric, goal-conditioned, and long-horizon manipulation, respectively.Source: §1 · p. 3

2 Results

The section frames general continual learning as organizing online, uncertain experience by separating conflicting distributions and integrating compatible ones. FlyGCL applies a biologically inspired hierarchy of expert specialization and temporal integration across perceptual and embodied tasks.

2 Results

  • The section targets online, uncertain, evolving streams, where the central challenge is separating conflicting distributions to reduce interference while integrating compatible ones to promote generalization.The biological account links spatially differentiated compartments with temporal memory scales.Source: §2 · p. 3, §2.1 · p. 6

2.1 Biological and computational basis of hierarchical modular framework

  • FlyGCL keeps the pretrained backbone as a stable representational substrate and uses random expansion for expert routing, diversified adaptive modules, and prediction heads with different effective memory windows.The hierarchy coordinates specialization across distributions with integration across spatial and temporal components.Source: §2.1 · p. 6
  • In the controlled olfactory model, MoE+EL reaches AUC of 43.2% versus the baseline's 28.9% under disjoint streams, and 30.1% versus 19.1% under blurry streams.The combined model also outperforms either component alone under joint streams.Source: Figure 2 · p. 5, §2.1 · p. 7
  • Experts are most accurate in their corresponding spatial regions, while under blurry streams MoE with different-rate temporal heads reaches anytime AUC of 30.1%, versus 27.4% with one head.Three equal-rate heads reach 28.4%; the progression holds with and without MoE in disjoint and blurry settings.Source: Figure 2 · p. 5, §2.1 · p. 7

2.2 General Continual Learning for Visual and Vision-Language Perception

  • In the primary comparison, FlyGCL’s Aauc/Alast exceed the strongest baseline by 3.5%/6.9% on CIFAR-100, 13.8%/18.7% on ImageNet-R, and 12.6%/24.6% on CUB-200, respectively.The performance gains are particularly clear on ImageNet-R and CUB-200.Source: §2.2 · p. 8
  • On CIFAR-100, FlyGCL reaches Aauc/Alast of 85.2/79.6%, and on ImageNet-R reaches 84.7/79.3%, with favorable forgetting and backward transfer.It achieves the strongest overall performance across both vision-language benchmarks.Source: §2.2 · p. 9, Figure 4 · p. 10

2.3 General Continual Learning for Embodied Perception and Action

  • FlyGCL reaches Aauc/Alast of 81.3%/82.6% on EgoExoLearn skill assessment and 62.1%/62.7% on EgoExo-Fitness, versus 58.0%/57.9% for the strongest competitor on the latter.Additional evaluations cover sequence verification and guidance-based execution verification.Source: Figure 5 · p. 11, §2.3 · p. 12
  • 83.1%/81.6% Alast/Aauc on LIBERO-Spatial and 86.1%/86.0% on LIBERO-Object exceed the strongest replay-free baselines by 49.5%/37.5% and 49.8%/35.9%, respectively.The same pattern holds for goal-conditioned and long-horizon manipulation.Source: Figure 6 · p. 13, §2.3 · p. 14

3 Discussion

FlyGCL organizes learning from dynamic experience through hierarchical coordination of expert specialization and ensemble integration. The discussion connects this approach to general continual learning and biological hypotheses while identifying boundaries for future work.

3 Discussion

  • Hierarchical specialization and integration organize learning according to relationships among evolving experiences, separating conflicting experience while integrating compatible experience.The framework is presented as a unified principle for general continual learning under online, uncertain, and evolving streams.Source: §3 · pp. 14–15
  • FlyGCL frames general continual learning as a route from static-data intelligence toward systems that develop through experience in long-running settings.Examples include embodied robots, autonomous vehicles, personalized assistants, and healthcare or scientific monitoring systems.Source: §3 · p. 15
  • Controlled olfactory experiments show that hierarchical coordination improves learning as experience shifts from disjoint to increasingly recurrent distributions.The model links sparse expansion and differentiated memory pathways to specialization and integration, offering a functional interpretation of the circuit organization.Source: §3 · p. 15
  • The study presents a bidirectional NeuroAI framework: Drosophila organization motivates machine-learning principles, and computational models generate testable neuroscience hypotheses.Predictions include reduced interference between dissimilar experiences and complementary contributions from memory pathways across spatial and temporal scales.Source: §3 · pp. 15–16
  • The study abstracts Drosophila circuitry, tests bounded online multimodal and embodied benchmarks, and concentrates plasticity in lightweight modules over stable pretrained representations.Richer biological mechanisms, open-ended interaction, and deeper learning within foundation models remain future directions.Source: §3 · p. 16

4 Methods

The methods formalize general continual learning as online learning over evolving streams, then analyze how modular predictors can separate conflicting experience and integrate compatible knowledge. FlyGCL implements this hierarchy over pretrained models, and the experiments specify evaluations across four continual-learning scenarios.

4.1 Problem Formulation

  • GCL models single-pass, non-stationary sessions with uncertain boundaries and asks modular predictors to retain knowledge while separating conflicting and sharing compatible distributions.Source: §4.1 · pp. 16–17

4.2 Theoretical Analysis

  • Expected GCL risk is bounded by observed-stream fitting error plus separation error, integration error, and coordination cost.Source: §4.2 · p. 18
  • Hierarchical modularity improves the risk bound when routing and aggregation gains exceed coordination cost.Source: §4.2 · p. 19
  • Pretrained representations can reduce coordination cost by making imperfect routing and integration less harmful in a stable semantic space.Source: §4.2 · pp. 19–20

4.3 FlyGCL Model

  • FlyGCL keeps the pretrained backbone stable and uses fixed random expansion for instance-level routing to specialized experts, not final prediction.Source: §4.3 · pp. 20–21
  • Each routed expert combines online and EMA heads at distinct timescales; visual recognition additionally calibrates temporal class probabilities with analytic evidence.Source: §4.3 · p. 22

4.4 Experimental Setups

  • Four scenarios are evaluated using final and anytime performance, with scenario-specific baselines and matched online budgets and trainable-module counts where applicable.Source: §4.4 · pp. 23–24

A.1 Proof of Theorem 1

The proof attributes the residual gap between expected stream risk and empirical fitting risk to separation, integration, and coordination costs, then bounds it by their sum under an additive excess-risk assumption.

A.1 Proof of Theorem 1

  • The residual gap between expected stream risk and empirical fitting risk is decomposed into separation, integration, and coordination excess risks.Insufficient separation of conflicting distributions creates Esep(F); isolating compatible distributions creates Eint(F); imperfect coordination contributes a third cost.Source: §A.1 · p. 35
  • Under the assumed additive excess-risk decomposition, the residual gap is upper bounded by the sum of the three non-negative components.Source: §A.1 · p. 35

A.2 Proof of Proposition 1

The proposition bounds hierarchical modular learning's expected GCL risk by fitting risk plus separation, integration, and coordination costs, and compares this bound with one without routing and integration gains.

A.2 Proof of Proposition 1

  • Hierarchical modular expected GCL risk is bounded by fitting risk plus separation, integration, and coordination costs.Source: §A.2 · p. 35
  • The proof bounds hierarchical separation error by the routed modular predictor's separation error and hierarchical integration error by the ensemble predictor's integration error.Source: §A.2 · p. 36
  • The bound comparison states that hierarchical modular learning improves the bound when its stated condition holds.The supplied passage does not specify that condition.Source: §A.2 · p. 36

A.3 Proof of Proposition 2

The proof bounds coordination cost through routing and aggregation mismatch, and argues that pretrained representations reduce the routing mismatch penalty and thus its upper bound.

A.3 Proof of Proposition 2

  • Coordination cost captures excess loss from routing or aggregation that mismatches the ideal stream-level organization.Source: §A.3 · p. 36
  • The total coordination cost is bounded by the sum of routing-induced and aggregation-induced mismatch.Routing error concerns assignment to a suboptimal module; aggregation mismatch concerns combining modular predictions that are not perfectly compatible.Source: §A.3 · p. 36
  • A pretrained backbone's more stable, semantically organized representations reduce the routing mismatch penalty and the upper bound on coordination cost.Source: §A.3 · p. 37

B Additional Experimental Setups

The additional setups describe a fixed, biologically grounded sensory simulation and online protocols for constructing and evaluating continual streams across benchmarks.

B.1 Biological Simulation

  • Disjoint, Blurry, and Joint odor streams vary boundary overlap while keeping each input unique and the total learning budget unchanged.Streams traverse five regions; Blurry retains 10 disjoint classes per region, while Joint applies overlap to all classes.Source: §B.1 · p. 37
  • The olfactory simulation uses a fixed sensory pathway with sparse PN-to-KC projections and retains the top 5% of KC responses for each odor.Training adds Gaussian noise to ORN responses, while evaluation is noise-free; learning occurs downstream of the fixed sensory encoding.Source: §B.1 · pp. 37–38
  • The odor MoE baseline uses five experts, routes by cosine similarity to observed stage-mean KC representations, and does not use labels at inference.Stage-to-expert association is specific to this protocol, not a requirement for known task identities or predefined stage semantics.Source: §B.1 · p. 38
  • The biological simulation compares Baseline, EL, MoE, and MoE+EL configurations under shared online observations, supervision, and replay-free learning budgets.Results use five independent datasets, streams, and initializations, with exposed-class accuracy evaluated at 25 checkpoints.Source: §B.1 · p. 38

B.2 Continual Stream Construction

  • Across benchmarks, stream sessions use prediction-aligned semantic units, and redistribution stays within official data splits.The table describes each benchmark's partition unit, task metric, and continual-learning metrics.Source: §B.2 · pp. 38–39, Table S1 · p. 39
  • The GCL protocol allows recurring samples outside their home sessions, whereas offline LIBERO presents ten tasks sequentially with multiple passes per task.Source: §B.2 · p. 39

C Additional Evaluation Metrics

The section defines task-specific metrics as session-level performance measures used in continual-learning aggregates. It also describes additional measures of forgetting and transfer for selected benchmarks.

C Additional Evaluation Metrics

  • Session-level performance R_t,j is instantiated with task-specific metrics in the continual-learning aggregates.The evaluation set for session j contains observations and their associated learning signals, and predictions are measured after learning through session t.Source: C Additional Evaluation Metrics · p. 39

C.1 Task Metrics

  • Classification accuracy evaluates categorical labels in visual recognition, vision-language classification, and action classification.Ranking-based mAP is also specified, and guidance-based execution verification reports accuracy and F1.Source: §C.1 · pp. 39–40
  • Skill assessment uses ranking accuracy, while action anticipation reports class-mean Top-5 recall separately for verb and noun prediction.Skill labels specify the relative quality ordering of two video observations.Source: §C.1 · p. 40
  • Binary sequence verification reports ROC-AUC and mAP for predicted verification scores.Guidance-based execution verification additionally reports accuracy and F1.Source: §C.1 · p. 40
  • Embodied vision-language-action learning is evaluated by the rate of successful task completion.The evaluation target is whether the learned policy completes an episode successfully.Source: §C.1 · pp. 40–41

C.2 Continual Learning Metrics

  • Selected benchmarks additionally report average forgetting, forward transfer, and negative backward transfer.Forgetting measures degradation from each session’s best historical performance to its final performance; higher forward transfer indicates stronger transfer to future sessions, while lower negative backward transfer indicates less interference.Source: §C.2 · p. 41

D Additional Implementation Details

Implementation details specify the pretrained components, the modules updated by FlyGCL, and shared evaluation protocols across methods.

D Additional Implementation Details

  • Visual recognition uses frozen ViT-B/16 backbones with supervised or self-supervised ImageNet pretraining, while vision-language learning uses frozen OpenAI CLIP ViT-B/16 encoders.Ego-exo video experiments retain benchmark-specific pretrained feature extraction and task heads, changing only the continual adaptation module.Source: D Additional Implementation Details · p. 42
  • Embodied experiments initialize a DiT flow-matching policy from a LIBERO-90 checkpoint built on DINOv2 vision and CLIP text encoders.Methods within each benchmark share the observation and policy pipeline.Source: D Additional Implementation Details · p. 42
  • FlyGCL freezes the pretrained backbone and updates only the active adaptive expert and output heads, with router updates from accumulated feature covariance and mean.Within each benchmark and task, methods share the stream, session order, preprocessing, mini-batch schedule, and evaluation checkpoints; replay memory is omitted unless intrinsically required by a method.Source: D Additional Implementation Details · p. 42

E Extensions beyond the Conference Version

The expanded work develops FlyGCL from a preliminary prompt-based visual-learning framework into a hierarchical principle with broader biological, theoretical, and empirical scope.

E Extensions beyond the Conference Version

  • FlyGCL frames general continual learning as coordinating specialization to separate conflicting experience and integration to promote generalization.It relates these functions to mixture-of-experts and ensemble learning under online, uncertain, evolving distributions.Source: E Extensions beyond the Conference… · p. 42
  • The biological account expands to spatially differentiated pathways and distinct temporal scales, tested in a controlled olfactory-learning model under varying distributional recurrence.The model evaluates spatial specialization, temporal integration, and their hierarchical coordination.Source: E Extensions beyond the Conference… · p. 42
  • The framework extends beyond prompts to adapters, LoRA, video modules, task heads, policy adapters, and action heads, with a new risk analysis of specialization, integration, and coordination.The analysis examines when routing and aggregation gains outweigh coordination costs and how stable pretrained representations can reduce the penalty of imperfect coordination.Source: E Extensions beyond the Conference… · p. 43
  • Visual experiments broaden evaluation across pretrained representations and prompt, adapter, and LoRA interfaces, including analysis of specialization and integration contributions.The experiments examine generality beyond a particular prompt design or backbone initialization.Source: E Extensions beyond the Conference… · p. 43
  • Evaluation extends to CLIP-based vision-language learning, ego-exo video understanding, and language-conditioned robotic manipulation under evolving task distributions.Together with the olfactory simulation, these settings broaden scope from visual classification to biological modeling, multimodal perception, temporal multi-view understanding, and embodied action.Source: E Extensions beyond the Conference… · p. 43

F Additional Results

Additional tables document performance comparisons and an ablation across visual recognition, vision-language, ego-exo, and embodied continual-learning benchmarks; the supplied passages give metric scope but not result values.

F Additional Results

  • Visual-recognition comparisons report anytime and final accuracy plus forgetting, and the ablation adds negative backward transfer.The comparisons use ViT-B/16 backbones with different pretraining; the ablation specifies ImageNet-21K pretraining and ImageNet-1K fine-tuning.Source: Table S3 · p. 45, Table S4 · p. 46
  • Vision-language benchmark comparisons report final and anytime accuracy, forgetting, and backward transfer.The caption specifies the GCL setting.Source: Table S5 · p. 46
  • Ego-exo tables cover skill-ranking accuracy, action Top-5 recall for verb and noun prediction, and action-classification accuracy.The skill-assessment and action-classification tables also report forgetting and anytime performance; the anticipation table reports final and anytime recall.Source: Table S6 · p. 47, Table S7 · p. 47, Table S8 · p. 47
  • EgoExo-Fitness sequence verification reports ROC-AUC and mAP, while guidance-based verification reports accuracy and F1.Both tables include final performance, forgetting, and anytime performance.Source: Table S10 · p. 48, Table S11 · p. 49
  • LIBERO embodied-learning tables report success, anytime success, forward transfer, and negative backward transfer under GCL and offline CL settings.The GCL and offline comparisons are reported separately.Source: Table S12 · p. 49, Table S13 · p. 50
Loading 2609.25146v1…