MOSAIC: Modality-Specific Adaptation for Incremental Continual Learning in Parkinson's Disease Gait Assessment

Martin J. McKeown; Minlin Zeng; Yang Qiu; Zhipeng Zhou; Zhiqi Shen

arxiv: 2606.13258 · v2 · pith:Z4LECRGZ · submitted 2026-06-11 · cs.AI

MOSAIC: Modality-Specific Adaptation for Incremental Continual Learning in Parkinson's Disease Gait Assessment

Minlin Zeng , Zhipeng Zhou , Yang Qiu , Martin J. McKeown , Zhiqi Shen This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-27 06:47 UTCgrok-4.3pith:Z4LECRGZ record.json open to challenge →

classification cs.AI

keywords continual learningParkinson's diseasegait assessmentmultimodalincremental learningmodality adaptationcatastrophic forgetting

0 comments

The pith

MOSAIC enables incremental addition of gait sensors for Parkinson's assessment by fixing toxic distillation, statistical shifts, and lost plasticity.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes a continual learning method for Parkinson's gait assessment where new sensor modalities arrive over time but old patient data cannot be reused due to privacy rules. It isolates three specific obstacles—unreliable cross-modal distillation from prior models, sensor-specific distribution shifts, and shrinking capacity to learn after old knowledge is protected—and counters each with a dedicated component inside the MOSAIC framework. Experiments across three multimodal gait datasets show higher final accuracy and less forgetting than prior continual-learning baselines. A reader would care because clinical systems routinely upgrade devices or combine centers, creating exactly this modality-incremental regime without the luxury of full retraining.

Core claim

MOSAIC identifies a Toxic Teacher effect in which distillation from an old model destabilizes new-modality representations, then counters it with a modality-specific warm-up phase, a statistics-decoupled MSBN backbone that keeps sensor statistics separate while sharing semantics, and a curriculum-guided repulsive objective that restores plasticity without erasing prior knowledge; the resulting system raises final performance and lowers forgetting on incremental multimodal Parkinson's gait tasks.

What carries the argument

The three-part MOSAIC mechanism: Modality-Specific Warm-Up before distillation, statistics-decoupled MSBN architecture, and curriculum-guided repulsive objective for plasticity recovery.

Load-bearing premise

The fixes for the toxic teacher, statistical mismatch, and plasticity loss will continue to work when entirely new sensor types appear in real clinical deployments outside the three tested datasets.

What would settle it

Run the same incremental protocol on a held-out fourth Parkinson's gait dataset that introduces a sensor type absent from the original three and check whether accuracy gains and forgetting reduction remain comparable to the reported results.

Figures

Figures reproduced from arXiv: 2606.13258 by Martin J. McKeown, Minlin Zeng, Yang Qiu, Zhipeng Zhou, Zhiqi Shen.

**Figure 1.** Figure 1: Illustration of the ”Teacher Misalignment” paradox in direct cross-modal KD. (a) Direct KD yields a pseudo-teacher with toxic, high-entropy outputs, whereas our Stage 1 Modality-Specific WarmUp rapidly stabilizes the entropy into a legitimate anchor. (b) Forcing alignment with a high-entropy teacher causes catastrophic negative transfer, which our method successfully prevents. data logistics challenge. In… view at source ↗

**Figure 2.** Figure 2: Overview of the proposed modality-incremental continual learning pipeline for Parkinson’s disease gait assessment. New sensing modalities are integrated sequentially without revisiting previous raw data. The framework combines modality-specific batch normalization to isolate variancesensitive statistics, EWC-style regularization and distillation to preserve previously learned semantics, and a repulsive ob… view at source ↗

**Figure 3.** Figure 3: Motivation analysis of shared BN under balanced joint training. Heterogeneous gait modalities are concatenated into shared BN, exposing mixed-statistics contamination and negative gradient interference in the shared convolutional backbone. adopt a three-stage optimization trajectory (summarized in Algorithm 1). 1) Stage 1: Modality-Specific Warm-Up: When transitioning between physically disjoint domains, … view at source ↗

**Figure 4.** Figure 4: Sensitivity analysis of the repulsive margin m. The selected setting (m = 0.3) yields the highest normalized average accuracy while maintaining favorable backward transfer, indicating a robust tradeoff between modality separation and knowledge retention. 0.1 0.3 0.5 1.0 3.0 5.0 8.0 −5 −4 −3 −2 −1 Optimal Curriculum Exponent (γ) BWT (%) BWT ↑ 0.97 0.98 0.99 1 NAA NAA ↑ [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗

**Figure 5.** Figure 5: Sensitivity analysis of the curriculum exponent γ. The selected setting (γ = 5.0) provides the best balance between backward transfer and normalized average accuracy, supporting a gradual schedule for introducing repulsive separation after semantic stabilization. E. Hyperparameters Analysis 1) Topological Margin (mT ) [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

read the original abstract

Gait-based Parkinson's disease assessment increasingly relies on heterogeneous sensors, but clinical systems rarely collect all modalities simultaneously. New sensors may arrive through device upgrades, protocol changes, or multi-center deployment, while historical patient data are often unavailable because of privacy and storage constraints. This modality-incremental setting faces three challenges: unreliable cross-modal distillation, modality-specific statistical shifts, and reduced plasticity after preservation. We propose MOSAIC, a compact continual learning framework. First, we identify the Toxic Teacher phenomenon and introduce Modality-Specific Warm-Up to stabilize newly learned modality representations before distillation. Second, we propose a statistics-decoupled MSBN architecture that isolates sensor statistics while maintaining a shared semantic backbone. Third, we design a curriculum-guided repulsive objective for Plasticity Recovery, preserving legacy knowledge while recovering modality-specific capacity. Experiments on three multimodal Parkinson's gait datasets show that MOSAIC improves final performance and mitigates forgetting. Project code is available at: https://github.com/minlinzeng/MOSAIC_Modality-Specific-Adaptation-for-Incremental-Continual-Learning-in-PD-Gait-Assessment.git

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MOSAIC gives a practical, code-backed tweak for adding new gait sensors over time in PD assessment without old data, but the abstract leaves the size of the gains unclear.

read the letter

The core takeaway is a compact framework for modality-incremental continual learning on Parkinson's gait data. New sensors arrive without historical records, and the work targets three specific headaches: bad cross-modal distillation, shifting sensor statistics, and loss of plasticity.

What is new is the named combination of Modality-Specific Warm-Up to handle the Toxic Teacher problem, a statistics-decoupled MSBN backbone, and a curriculum-guided repulsive loss for plasticity recovery. These are tailored extensions rather than a wholesale new theory, but they fit the clinical constraint of incremental sensor changes. Releasing the code is a clear plus for anyone who wants to check the implementation.

The experiments claim better final accuracy and reduced forgetting across three multimodal PD gait datasets. That matches the practical need the abstract describes.

The main soft spot is the lack of any numbers, baselines, error bars, or ablation results in the abstract. Without those, it is difficult to judge whether the gains are large enough or stable enough to matter in deployment. The generalization claim beyond the three datasets also rests on the assumption that the fixes will not create new instabilities when fresh modalities appear.

This paper is aimed at researchers working on continual learning for medical sensor streams, especially in neurology or gait analysis. A reader already focused on modality shifts in healthcare will get the most out of it.

It deserves a serious referee. The problem is real, the approach is coherent, and the code link allows direct checking. I would send it for review and ask the authors to supply the missing quantitative details.

Referee Report

0 major / 1 minor

Summary. The manuscript proposes MOSAIC, a compact continual learning framework for the modality-incremental setting in Parkinson's disease gait assessment. It identifies the Toxic Teacher phenomenon and introduces three components: Modality-Specific Warm-Up to stabilize new modality representations before distillation, a statistics-decoupled MSBN architecture to isolate sensor statistics while sharing a semantic backbone, and a curriculum-guided repulsive objective for plasticity recovery. The central claim, supported by experiments on three multimodal Parkinson's gait datasets, is that MOSAIC improves final performance and mitigates forgetting; project code is provided.

Significance. If the results hold, the work addresses a practically relevant problem in clinical multimodal sensing where new modalities arrive over time without access to prior patient data due to privacy constraints. The explicit identification of modality-specific challenges and the release of code for reproducibility are strengths that could support adoption in medical continual-learning applications.

minor comments (1)

[Abstract] Abstract: the claim of improved performance is stated without any quantitative metrics, baselines, dataset sizes, or error bars; including at least the key performance numbers and a brief comparison table would strengthen the summary.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary of our work on MOSAIC for modality-incremental continual learning in Parkinson's disease gait assessment and for recommending minor revision. The report correctly captures the identification of the Toxic Teacher phenomenon and the three proposed components. No major comments are provided in the report.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents an empirical continual learning framework (MOSAIC) consisting of three proposed components—Modality-Specific Warm-Up, statistics-decoupled MSBN, and curriculum-guided repulsive objective—validated through experiments on three multimodal Parkinson's gait datasets. No mathematical derivations, equations, or predictions are described that reduce to fitted parameters by construction, self-definitional loops, or load-bearing self-citations. The central claims rest on experimental results rather than any internal reduction of outputs to inputs, making the derivation chain self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the method relies on standard deep learning training assumptions.

axioms (1)

standard math Standard assumptions of neural network optimization and continual learning hold for the proposed components.
The framework description presupposes typical DL training dynamics without stating exceptions.

pith-pipeline@v0.9.1-grok · 5733 in / 1152 out tokens · 19964 ms · 2026-06-27T06:47:34.637845+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 5 canonical work pages · 2 internal anchors

[1]

Tackling parkinson’s disease as a global challenge,

R. Bhidayasiri, L. V . Kalia, and B. R. Bloem, “Tackling parkinson’s disease as a global challenge,” pp. 1277–1280, 2023

2023
[2]

The parkinson pandemic—a call to action,

E. R. Dorsey and B. R. Bloem, “The parkinson pandemic—a call to action,”JAMA neurology, vol. 75, no. 1, pp. 9–10, 2018

2018
[3]

Towards re- laxed multimodal inputs for gait-based parkinson’s disease assessment,

M. Zeng, Z. Zhou, Y . Qiu, M. J. McKeown, and Z. Shen, “Towards re- laxed multimodal inputs for gait-based parkinson’s disease assessment,” arXiv preprint arXiv:2510.15748, 2025

work page arXiv 2025
[4]

A survey of artificial intelligence in gait-based neurodegenerative disease diagnosis,

H. Rao, M. Zeng, X. Zhao, and C. Miao, “A survey of artificial intelligence in gait-based neurodegenerative disease diagnosis,”Neuro- computing, vol. 626, p. 129533, 2025

2025
[5]

Harmony: A unified framework for modality incremental learning,

Y . Song, X. Yang, D. Jiang, Y . Wang, and C. Xu, “Harmony: A unified framework for modality incremental learning,”arXiv preprint arXiv:2504.13218, 2025

work page arXiv 2025
[6]

Breaking data silos in parkinson’s disease diagnosis: An adaptive federated learning approach for privacy-preserving facial expression analysis,

M. Pang, H. Xu, Z. Huang, Y . Zhou, W. Huang, and B. Wang, “Breaking data silos in parkinson’s disease diagnosis: An adaptive federated learning approach for privacy-preserving facial expression analysis,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 13, 2025, pp. 14 352–14 360

2025
[7]

Continual self-supervised learning: Towards universal multi-modal medical data representation learning,

Y . Ye, Y . Xie, J. Zhang, Z. Chen, Q. Wu, and Y . Xia, “Continual self-supervised learning: Towards universal multi-modal medical data representation learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 11 114–11 124

2024
[8]

Confusion mixup regularized multimodal fusion network for contin- ual egocentric activity recognition,

H. Wang, S. Zhou, Q. Wu, H. Li, F. Meng, L. Xu, and H. Qiu, “Confusion mixup regularized multimodal fusion network for contin- ual egocentric activity recognition,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3560–3569

2023
[9]

Modality- incremental learning with disjoint relevance mapping networks for image-based semantic segmentation,

N. Hegde, S. Muralidhara, R. Schuster, and D. Stricker, “Modality- incremental learning with disjoint relevance mapping networks for image-based semantic segmentation,” in2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV). IEEE, 2025, pp. 5540–5549

2025
[10]

Learning without isolation: Pathway protection for continual learning,

Z. Chen, A. Wuerkaixi, S. Cui, H. Li, D. Li, J. Zhang, B. Han, G. Niu, H. Liu, Y . Yanget al., “Learning without isolation: Pathway protection for continual learning,” inForty-second International Conference on Machine Learning, 2025

2025
[11]

Contrastive regularization over lora for multimodal biomedical image incremental learning,

H. Zhang, Y . Liang, H. Kuang, L. Cen, Z. Qu, Y . Cen, M. Zeng, and S. Kan, “Contrastive regularization over lora for multimodal biomedical image incremental learning,” inProceedings of the 33rd ACM Interna- tional Conference on Multimedia, 2025, pp. 7500–7509

2025
[12]

Continual retinal vision- language pre-training upon incremental imaging modalities,

Y . Yao, R. Wu, Y . Zhou, and T. Zhou, “Continual retinal vision- language pre-training upon incremental imaging modalities,” inInterna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 111–121

2025
[13]

Harmi: Human activity recognition via multi-modality incremental learning,

X. Zhang, H. Yu, Y . Yang, J. Gu, Y . Li, F. Zhuang, D. Yu, and Z. Ren, “Harmi: Human activity recognition via multi-modality incremental learning,”IEEE Journal of Biomedical and Health Informatics, vol. 26, no. 3, pp. 939–951, 2021

2021
[14]

Artificial intelligence- enabled detection and assessment of parkinson’s disease using multi- modal data: A survey,

A. Zhao, Y . Liu, X. Yu, X. Xing, and H. Zhou, “Artificial intelligence- enabled detection and assessment of parkinson’s disease using multi- modal data: A survey,”Information Fusion, p. 103175, 2025

2025
[15]

Multimodal gait recognition for neurodegenerative diseases,

A. Zhao, J. Li, J. Dong, L. Qi, Q. Zhang, N. Li, X. Wang, and H. Zhou, “Multimodal gait recognition for neurodegenerative diseases,”IEEE transactions on cybernetics, vol. 52, no. 9, pp. 9439–9453, 2021

2021
[16]

Graph fusion network-based multimodal learning for freezing of gait detection,

K. Hu, Z. Wang, K. A. E. Martens, M. Hagenbuchner, M. Bennamoun, A. C. Tsoi, and S. J. Lewis, “Graph fusion network-based multimodal learning for freezing of gait detection,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 3, pp. 1588–1600, 2021

2021
[17]

Multi-modal gait recognition via effective spatial- temporal feature fusion,

Y . Cui and Y . Kang, “Multi-modal gait recognition via effective spatial- temporal feature fusion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 17 949–17 957

2023
[18]

Exploring tradeoffs through mode connectivity for multi-task learning,

Z. Zhou, Z. Meng, P. Wu, P. Zhao, and C. Miao, “Exploring tradeoffs through mode connectivity for multi-task learning,”Advances in Neural Information Processing Systems, vol. 38, pp. 81 268–81 285, 2026

2026
[19]

Injecting imbalance sensitivity for multi-task learning,

Z. Zhou, L. Liu, P. Zhao, and W. Gong, “Injecting imbalance sensitivity for multi-task learning,”arXiv preprint arXiv:2503.08006, 2025

work page arXiv 2025
[20]

Continual Learning for VLMs: A Survey and Taxonomy Beyond Forgetting

Y . Liu, Q. Hong, L. Huang, A. Gomez-Villa, D. Goswami, X. Liu, J. van de Weijer, and Y . Tian, “Continual learning for vlms: A survey and taxonomy beyond forgetting,”arXiv preprint arXiv:2508.04227, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[21]

Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning,

V . W. Liang, Y . Zhang, Y . Kwon, S. Yeung, and J. Y . Zou, “Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning,”Advances in Neural Information Processing Systems, vol. 35, pp. 17 612–17 625, 2022

2022
[22]

Continual optimization with symmetry teleportation for multi-task learning,

Z. Zhou, Z. Meng, P. Wu, P. Zhao, and C. Miao, “Continual optimization with symmetry teleportation for multi-task learning,”Advances in Neural Information Processing Systems, vol. 38, pp. 171 209–171 230, 2026

2026
[23]

Delve into the Applicability of Advanced Optimizers for Multi-Task Learning

Z. Zhou, L. Cao, P. Wu, P. Zhao, and C. Miao, “Delve into the applicability of advanced optimizers for multi-task learning,”arXiv preprint arXiv:2604.08939, 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026
[24]

Batch normalization: Accelerating deep network training by reducing internal covariate shift,

S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inInternational conference on machine learning. pmlr, 2015, pp. 448–456

2015
[25]

Overcoming catastrophic forgetting in neural networks,

J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,”Pro- ceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017

2017
[26]

Learning without forgetting,

Z. Li and D. Hoiem, “Learning without forgetting,”IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 12, pp. 2935– 2947, 2017

2017
[27]

Gradient episodic memory for continual learning,

D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,”Advances in neural information processing systems, vol. 30, 2017

2017
[28]

Mea- suring catastrophic forgetting in neural networks,

R. Kemker, M. McClure, A. Abitino, T. Hayes, and C. Kanan, “Mea- suring catastrophic forgetting in neural networks,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

2018
[29]

Wearables for gait in parkinson’s disease and age- matched controls,

K. Kontsonet al., “Wearables for gait in parkinson’s disease and age- matched controls,”Synapse by SAGE Bionetworks, 2024

2024
[30]

A public data set of walk- ing full-body kinematics and kinetics in individuals with parkinson’s disease,

T. K. F. Shida, T. M. Costa, C. E. N. de Oliveira, R. de Castro Treza, S. M. Hondo, E. Los Angeles, C. Bernardo, L. dos Santos de Oliveira, M. de Jesus Carvalho, and D. B. Coelho, “A public data set of walk- ing full-body kinematics and kinetics in individuals with parkinson’s disease,”Frontiers in Neuroscience, vol. 17, p. 992585, 2023

2023
[31]

A public data set of videos, inertial measurement unit, and clinical scales of freezing of gait in individuals with parkinson’s disease during a turning-in-place task,

C. Ribeiro De Souza, R. Miao, J. ´Avila De Oliveira, A. Cristina De Lima-Pardini, D. Fragoso De Campos, C. Silva-Batista, L. Teixeira, S. Shokur, B. Mohamed, and D. B. Coelho, “A public data set of videos, inertial measurement unit, and clinical scales of freezing of gait in individuals with parkinson’s disease during a turning-in-place task,” Frontiers i...

2022
[32]

Order-robust class incremental learning: Graph-driven dynamic similarity grouping,

G. Lai, Y . Li, X. Wang, J. Zhang, T. Li, and X. Yang, “Order-robust class incremental learning: Graph-driven dynamic similarity grouping,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 4894–4904

2025

[1] [1]

Tackling parkinson’s disease as a global challenge,

R. Bhidayasiri, L. V . Kalia, and B. R. Bloem, “Tackling parkinson’s disease as a global challenge,” pp. 1277–1280, 2023

2023

[2] [2]

The parkinson pandemic—a call to action,

E. R. Dorsey and B. R. Bloem, “The parkinson pandemic—a call to action,”JAMA neurology, vol. 75, no. 1, pp. 9–10, 2018

2018

[3] [3]

Towards re- laxed multimodal inputs for gait-based parkinson’s disease assessment,

M. Zeng, Z. Zhou, Y . Qiu, M. J. McKeown, and Z. Shen, “Towards re- laxed multimodal inputs for gait-based parkinson’s disease assessment,” arXiv preprint arXiv:2510.15748, 2025

work page arXiv 2025

[4] [4]

A survey of artificial intelligence in gait-based neurodegenerative disease diagnosis,

H. Rao, M. Zeng, X. Zhao, and C. Miao, “A survey of artificial intelligence in gait-based neurodegenerative disease diagnosis,”Neuro- computing, vol. 626, p. 129533, 2025

2025

[5] [5]

Harmony: A unified framework for modality incremental learning,

Y . Song, X. Yang, D. Jiang, Y . Wang, and C. Xu, “Harmony: A unified framework for modality incremental learning,”arXiv preprint arXiv:2504.13218, 2025

work page arXiv 2025

[6] [6]

Breaking data silos in parkinson’s disease diagnosis: An adaptive federated learning approach for privacy-preserving facial expression analysis,

M. Pang, H. Xu, Z. Huang, Y . Zhou, W. Huang, and B. Wang, “Breaking data silos in parkinson’s disease diagnosis: An adaptive federated learning approach for privacy-preserving facial expression analysis,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 13, 2025, pp. 14 352–14 360

2025

[7] [7]

Continual self-supervised learning: Towards universal multi-modal medical data representation learning,

Y . Ye, Y . Xie, J. Zhang, Z. Chen, Q. Wu, and Y . Xia, “Continual self-supervised learning: Towards universal multi-modal medical data representation learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 11 114–11 124

2024

[8] [8]

Confusion mixup regularized multimodal fusion network for contin- ual egocentric activity recognition,

H. Wang, S. Zhou, Q. Wu, H. Li, F. Meng, L. Xu, and H. Qiu, “Confusion mixup regularized multimodal fusion network for contin- ual egocentric activity recognition,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3560–3569

2023

[9] [9]

Modality- incremental learning with disjoint relevance mapping networks for image-based semantic segmentation,

N. Hegde, S. Muralidhara, R. Schuster, and D. Stricker, “Modality- incremental learning with disjoint relevance mapping networks for image-based semantic segmentation,” in2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV). IEEE, 2025, pp. 5540–5549

2025

[10] [10]

Learning without isolation: Pathway protection for continual learning,

Z. Chen, A. Wuerkaixi, S. Cui, H. Li, D. Li, J. Zhang, B. Han, G. Niu, H. Liu, Y . Yanget al., “Learning without isolation: Pathway protection for continual learning,” inForty-second International Conference on Machine Learning, 2025

2025

[11] [11]

Contrastive regularization over lora for multimodal biomedical image incremental learning,

H. Zhang, Y . Liang, H. Kuang, L. Cen, Z. Qu, Y . Cen, M. Zeng, and S. Kan, “Contrastive regularization over lora for multimodal biomedical image incremental learning,” inProceedings of the 33rd ACM Interna- tional Conference on Multimedia, 2025, pp. 7500–7509

2025

[12] [12]

Continual retinal vision- language pre-training upon incremental imaging modalities,

Y . Yao, R. Wu, Y . Zhou, and T. Zhou, “Continual retinal vision- language pre-training upon incremental imaging modalities,” inInterna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 111–121

2025

[13] [13]

Harmi: Human activity recognition via multi-modality incremental learning,

X. Zhang, H. Yu, Y . Yang, J. Gu, Y . Li, F. Zhuang, D. Yu, and Z. Ren, “Harmi: Human activity recognition via multi-modality incremental learning,”IEEE Journal of Biomedical and Health Informatics, vol. 26, no. 3, pp. 939–951, 2021

2021

[14] [14]

Artificial intelligence- enabled detection and assessment of parkinson’s disease using multi- modal data: A survey,

A. Zhao, Y . Liu, X. Yu, X. Xing, and H. Zhou, “Artificial intelligence- enabled detection and assessment of parkinson’s disease using multi- modal data: A survey,”Information Fusion, p. 103175, 2025

2025

[15] [15]

Multimodal gait recognition for neurodegenerative diseases,

A. Zhao, J. Li, J. Dong, L. Qi, Q. Zhang, N. Li, X. Wang, and H. Zhou, “Multimodal gait recognition for neurodegenerative diseases,”IEEE transactions on cybernetics, vol. 52, no. 9, pp. 9439–9453, 2021

2021

[16] [16]

Graph fusion network-based multimodal learning for freezing of gait detection,

K. Hu, Z. Wang, K. A. E. Martens, M. Hagenbuchner, M. Bennamoun, A. C. Tsoi, and S. J. Lewis, “Graph fusion network-based multimodal learning for freezing of gait detection,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 3, pp. 1588–1600, 2021

2021

[17] [17]

Multi-modal gait recognition via effective spatial- temporal feature fusion,

Y . Cui and Y . Kang, “Multi-modal gait recognition via effective spatial- temporal feature fusion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 17 949–17 957

2023

[18] [18]

Exploring tradeoffs through mode connectivity for multi-task learning,

Z. Zhou, Z. Meng, P. Wu, P. Zhao, and C. Miao, “Exploring tradeoffs through mode connectivity for multi-task learning,”Advances in Neural Information Processing Systems, vol. 38, pp. 81 268–81 285, 2026

2026

[19] [19]

Injecting imbalance sensitivity for multi-task learning,

Z. Zhou, L. Liu, P. Zhao, and W. Gong, “Injecting imbalance sensitivity for multi-task learning,”arXiv preprint arXiv:2503.08006, 2025

work page arXiv 2025

[20] [20]

Continual Learning for VLMs: A Survey and Taxonomy Beyond Forgetting

Y . Liu, Q. Hong, L. Huang, A. Gomez-Villa, D. Goswami, X. Liu, J. van de Weijer, and Y . Tian, “Continual learning for vlms: A survey and taxonomy beyond forgetting,”arXiv preprint arXiv:2508.04227, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025

[21] [21]

Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning,

V . W. Liang, Y . Zhang, Y . Kwon, S. Yeung, and J. Y . Zou, “Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning,”Advances in Neural Information Processing Systems, vol. 35, pp. 17 612–17 625, 2022

2022

[22] [22]

Continual optimization with symmetry teleportation for multi-task learning,

Z. Zhou, Z. Meng, P. Wu, P. Zhao, and C. Miao, “Continual optimization with symmetry teleportation for multi-task learning,”Advances in Neural Information Processing Systems, vol. 38, pp. 171 209–171 230, 2026

2026

[23] [23]

Delve into the Applicability of Advanced Optimizers for Multi-Task Learning

Z. Zhou, L. Cao, P. Wu, P. Zhao, and C. Miao, “Delve into the applicability of advanced optimizers for multi-task learning,”arXiv preprint arXiv:2604.08939, 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026

[24] [24]

Batch normalization: Accelerating deep network training by reducing internal covariate shift,

S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inInternational conference on machine learning. pmlr, 2015, pp. 448–456

2015

[25] [25]

Overcoming catastrophic forgetting in neural networks,

J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,”Pro- ceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017

2017

[26] [26]

Learning without forgetting,

Z. Li and D. Hoiem, “Learning without forgetting,”IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 12, pp. 2935– 2947, 2017

2017

[27] [27]

Gradient episodic memory for continual learning,

D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,”Advances in neural information processing systems, vol. 30, 2017

2017

[28] [28]

Mea- suring catastrophic forgetting in neural networks,

R. Kemker, M. McClure, A. Abitino, T. Hayes, and C. Kanan, “Mea- suring catastrophic forgetting in neural networks,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

2018

[29] [29]

Wearables for gait in parkinson’s disease and age- matched controls,

K. Kontsonet al., “Wearables for gait in parkinson’s disease and age- matched controls,”Synapse by SAGE Bionetworks, 2024

2024

[30] [30]

A public data set of walk- ing full-body kinematics and kinetics in individuals with parkinson’s disease,

T. K. F. Shida, T. M. Costa, C. E. N. de Oliveira, R. de Castro Treza, S. M. Hondo, E. Los Angeles, C. Bernardo, L. dos Santos de Oliveira, M. de Jesus Carvalho, and D. B. Coelho, “A public data set of walk- ing full-body kinematics and kinetics in individuals with parkinson’s disease,”Frontiers in Neuroscience, vol. 17, p. 992585, 2023

2023

[31] [31]

A public data set of videos, inertial measurement unit, and clinical scales of freezing of gait in individuals with parkinson’s disease during a turning-in-place task,

C. Ribeiro De Souza, R. Miao, J. ´Avila De Oliveira, A. Cristina De Lima-Pardini, D. Fragoso De Campos, C. Silva-Batista, L. Teixeira, S. Shokur, B. Mohamed, and D. B. Coelho, “A public data set of videos, inertial measurement unit, and clinical scales of freezing of gait in individuals with parkinson’s disease during a turning-in-place task,” Frontiers i...

2022

[32] [32]

Order-robust class incremental learning: Graph-driven dynamic similarity grouping,

G. Lai, Y . Li, X. Wang, J. Zhang, T. Li, and X. Yang, “Order-robust class incremental learning: Graph-driven dynamic similarity grouping,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 4894–4904

2025