Web-connected sensing applications continuously generate multivariate time series with evolving behavioral categories, making Time Series Class-Incremental Learning (TSCIL) essential for adapting to new classes without forgetting previously acquired knowledge. However, preserving diverse and evolving class distributions under strict privacy and memory constraints remains a fundamental challenge. Existing approaches often face a trade-off between retaining sufficient historical knowledge for accurate recognition and minimizing the memory required for continual adaptation. To address these limitations, we propose TRACE, a memory-efficient, replay-free framework that preserves class-specific knowledge through generative modeling rather than sample rehearsal. TRACE employs a compact, frozen temporal-spectral representation and lightweight class-specific autoencoders to capture complex intra-class variations. By independently preserving each class model, TRACE avoids recursive generation and catastrophic forgetting, while enabling direct classification through reconstruction-based scoring without retaining historical samples or large statistical matrices. Extensive experiments across five TSCIL benchmarks demonstrate that TRACE achieves competitive classification accuracy while reducing mean retained memory by approximately 98.26% compared with TS-ACL, the strongest evaluated analytical baseline, and achieving an approximately 33.43% relative improvement in accuracy on GrabMyo.
Figures & tables
Figure 1. Mean final test accuracy versus retained memory across five datasets (25 runs per dataset). Bubble area represents the mean within-dataset variance across runs, and the dashed line indicates the offline joint-training reference. Accuracy versus retained state for the same methods as Table 2, equally averaged across five datasets. Ours uses a frozen CNN with class-specific deterministic AEs. Its point is at 9.1654 MiB and 88.8538 percent. Circle area is 4.9 times the mean of five within-dataset sample run variances, without a size floor, and does not represent a confidence interval. The accuracy axis spans 10 to 100 percent. Offline is shown as a dashed horizontal reference.
Figure 2. TRACE uses deterministic class-specific autoencoders. (a) Task 0 supervises the temporal-spectral CNN. The temporary head is discarded, and the CNN and Task-0 feature statistics are frozen. Each incoming class fits a bottleneck AE with reconstruction loss, then freezes it without replay. (b) Queries are encoded once and reconstructed by all seen-class AEs. The smallest feature MSE determines the prediction. Snowflakes mark frozen components. Two-panel \methodnameplain{} AE architecture. Training combines raw time series and rFFT magnitudes, supervises a four-block CNN on Task 0, and freezes the multi-scale representation and Task-0 standardization. Each class has a deterministic encoder, a 20-dimensional bottleneck and a decoder trained by reconstruction loss. No variance branch, sampling or KL term is used. Trained class AEs are frozen and stored. Testing compares deterministic reconstructions from all seen-class AEs with the same query feature and predicts the minimum-MSE class. Waveforms and score bars are schematic.
Dataset
Shape ( L×M )
Classes
Orig. Tasks
Our Tasks
HAR ( Anguita et al., 2013 )
128×9
6
3
3
UWave ( Liu et al., 2009 )
315×3
8
4
4
DailySports ( Altun and Barshan, 2010 )
125×45
18
6
9
WISDM ( Weiss et al., 2019 )
200×3
18
6
9
GrabMyo ( Pradhan et al., 2022 )
128×28
16
5
8
Table 1. Benchmark class counts, task horizons (2 classes per task), and input dimensions ( L×M ) across the five TSCIL benchmarks compared with the original benchmark ( Qiao et al., 2024 ) .
Table 2. Final test accuracy (mean ± SD, %) and estimated retained memory (MiB) over five class orders and five seeds per dataset (25 runs). Macro averages weight datasets equally. Offline denotes all-class joint training. Bold and underlining indicate the highest and second-highest five-dataset macro-average accuracy among continual-learning methods, respectively.
Figure 3. GrabMyo per-class test accuracy changes (pp) are measured from first evaluation and averaged over five seeds under a fixed class arrival sequence. Blue solid curves show Ours (AE), and orange dashed curves show GR. Sixteen panels plot class-wise changes in test accuracy over eight tasks on GrabMyo under a fixed class arrival sequence. Ours uses deterministic class-specific AEs on frozen temporal-spectral features and retains more old-class accuracy than GR under this sequence, while still showing some decline as new classes arrive. The two classes introduced at the final task have only their initial zero-change point.
Figure 4. Raw-series AE (top) and feature AE (ours, bottom) deterministically reconstruct identical validation inputs. Shading shows real density, and dots show paired reconstructions in a shared joint t-SNE diagnostic space per column. All classes are shown for HAR/UWave/DailySports. Six classes selected post hoc are shown for WISDM/GrabMyo. A two-row, five-column figure. Columns are HAR, UWave, DailySports, WISDM and GrabMyo. Deterministic raw-series AE reconstructions are above feature AE reconstructions. The first three columns show all 6, 8 and 18 classes, with 98, 10 and 36 paired validation inputs per class respectively. WISDM shows classes 0, 1, 2, 3, 4 and 13 with 96 inputs per class. GrabMyo shows classes 5, 6, 10, 13, 14 and 15 with 200 inputs per class. In every column the same real inputs, class colors, background density and joint t-SNE coordinates are shared by both rows. Each real input has exactly one deterministic reconstruction in each row. Raw reconstructions pass through the same frozen initial-task CNN and frozen all-class diagnostic MLP as the real inputs. Feature reconstructions pass through the same diagnostic MLP, which is used only for visualization. WISDM and GrabMyo classes were selected post hoc for neighborhood separation. All three other datasets use all classes.
Figure 5. GrabMyo per-class accuracy gains over N=0 (pp) average five seeds under one class arrival sequence (Task 0: c4, c14). Solid/dashed outlines indicate CNN-pretrained/unpretrained classes. Offline is the all-class linear head. Darker red indicates larger gains. Heatmap of final per-class AE accuracy gains relative to the N0 random frozen CNN plus AE baseline on GrabMyo under the class arrival sequence starting with Task 0 classes c4 and c14. Each cell averages five paired seeds. Rows N1, N2 and N4 retain deterministic per-class AEs. The separate Offline linear row uses the all-class CNN classifier head. The two panels preserve class arrival order. Solid outlines mark pretrained classes and the dashed region marks the fixed eight classes unpretrained under N up to 4. All 64 displayed mean gains are positive. A white-to-red sequential scale spans zero to 30 percentage points, with darker red indicating larger gains.
Variant
HAR
UWave
DailySp.
WISDM
GrabMyo
Feature representation (AE)
No Map
94.43±0.48
96.00±0.25
97.35±0.71
30.02±0.95
42.07±0.70
Last (128)
84.01±2.41
92.59±1.16
96.17±1.34
74.64±4.07
61.46±3.65
Base (576)
90.06±1.10
93.22±0.58
98.52±0.26
80.02±2.34
82.46±0.90
Wide (1152)
90.92±0.64
93.32±0.65
98.61±0.14
79.20±2.34
85.19±0.72
Input design (AE)
Table 3. Feature representation, input design, and class-model ablations report final test accuracy (mean ± SD, %, 25 runs per cell). Bold indicates the column best.
N / Head
HAR
UWave
DailySp.
GrabMyo
N=0
89.55±0.70
92.79±0.59
98.69±0.21
73.82±0.58
N=1
90.06±1.10
93.22±0.58
98.52±0.26
82.46±0.90
N=2
92.16±1.65
94.66±0.61
98.81±0.19
86.67±0.64
N=4
—
—
99.15±0.26
91.34±0.34
All AE
94.56±0.76
96.76±0.24
99.57±0.09
95.19±0.20
Linear
94.30±0.62
95.80±1.11
99.09±0.24
91.42±1.05
Table 4. CNN pretraining coverage is evaluated by test accuracy (mean ± SD, %, 25 runs per cell). N counts pretraining tasks. All AE uses all-class pretraining, and Linear is the corresponding CNN classifier.
Time series data play a pivotal role across numerous domains, including healthcare and manufacturing. In real-world environments, models must cope with distribution shifts over time, a challenge commonly addressed through Continual Learning (CL) techniques. However, existing CL methods face a critical limitation: real-world data streams are rarely fully labeled, making annotation cost a major practical constraint. This paper investigates Active Class-Incremental Learning (ACIL) for multivariate time series, where a model must sequentially learn new classes while selectively querying labels under a fixed annotation budget. We present a systematic evaluation of a wide range of query strategies combined with multiple rehearsal-based approaches, assessing their impact on plasticity, stability, and label efficiency across four benchmark datasets. Our analysis reveals the limitations of uncertainty-based and distribution-aware methods in achieving strong performance under constrained labeling budgets. To address these shortcomings, we propose TypiCore, a novel hybrid query strategy that alternates between typicality-based and diversity-based sample selection across active learning cycles, enabling the construction of memory buffers that are both representative and diverse. Evaluated on the TSCIL benchmark, TypiCore delivers statistically significant improvements over all baselines and matches or surpasses fully supervised continual learning performance on multiple datasets while requiring a fraction of the available labels.
Gabor Szucs, Samuel Jacsev, Marcell Nemeth +2
Department of Telecommunications and Artificial Intelligence, Budapest University of Technology and Economics, Budapest, Hungary. · Department of Information Engineering, University of Padua, Padua, Italy.
This work presents a generative continual learning framework based on growing self-organizing maps (GSOMs) that are augmented with learned distributional statistics as well as encoder-decoder models for class-incremental learning. The proposed approach enables exemplar-free replay using distributional statistical memory, which eliminates the need to store raw data. Each GSOM unit maintains its own mean, variance, and covariance estimates, which are subsequently used to generate synthetic samples for replay; in encoder-decoder configurations, these samples are then decoded back into the input space (via ancestral sampling) for subsequent training. Our method is fully unsupervised, as it does not rely on explicit task boundaries or class labels during training. Results across multiple benchmarks show that the proposed approach achieves performance competitive even with supervised state-of-the-art memory-based methods while consistently outperforming memory-free approaches. In several settings, our framework matches or exceeds existing baselines, particularly in challenging single-class incremental scenarios. We also provide baseline results for single-class incremental TinyImageNet and MiniImageNet, offering a useful reference for future work. This work highlights the effectiveness of an unsupervised, adaptive, topology-driven neural form of statistical replay as a scalable, flexible approach to continual learning.
Pujan Thapa, Alexander Ororbia, Travis Desell
Department of Computer Science Rochester Institute of Technology Rochester, NY, USA
Many systems used in real-world environments require adding new categories and incorporating new information without forgetting what was previously learnt by the classification model. This is known as class-incremental continual learning, and in the case of multivariate time-series, is further complicated by the temporal structure of the data. In this paper, we present a novel approach for performing class incremental continual learning for the classification of multivariate time series data based upon the construction of a dual-stream feature extraction pipeline (using both deep temporal embedding features generated via a pre-trained frozen foundation model and application of statistical features). Evaluated on five benchmark datasets, the proposed system achieves competitive average accuracy across all datasets while maintaining low forgetting rates across all experimental configurations.
Pablo García-Santaclara, Bruno Fernández-Castro, Rebeca Pilar Díaz-Redondo
atlanTTic – ICLAB, Universidade de Vigo, Vigo 36310, Spain · Centro Tecnolóxico de Telecomunicacións de Galicia (GRADIANT), Vigo 36214, Spain