AnchorInv: Few-Shot Class-Incremental Learning of Physiological Signals via Representation Space Guided Inversion

TL;DR

This work tackles few-shot class-incremental learning (FSCIL) for physiological time-series data under privacy constraints, where storing raw samples is undesirable. It introduces AnchorInv, a method that builds a feature-space anchor set from base-session embeddings and uses MAE-driven inversion to synthesize replay samples aligned to those anchors, enabling faithful retention of prior knowledge while adapting to new classes. Through extensive experiments on BCI, NHIE, and GRABMyo datasets, AnchorInv consistently outperforms state-of-the-art baselines in Macro-F1 across base and incremental classes, with statistically significant improvements and strong qualitative analysis of inverted samples. The approach demonstrates that guiding inversion with anchor points provides privacy-preserving, effective regularization for backbone finetuning in FSCIL, with practical guidance on anchor selection and shot counts for robust deployment.

Abstract

Deep learning models have demonstrated exceptional performance in a variety of real-world applications. These successes are often attributed to strong base models that can generalize to novel tasks with limited supporting data while keeping prior knowledge intact. However, these impressive results are based on the availability of a large amount of high-quality data, which is often lacking in specialized biomedical applications. In such fields, models are usually developed with limited data that arrive incrementally with novel categories. This requires the model to adapt to new information while preserving existing knowledge. Few-Shot Class-Incremental Learning (FSCIL) methods offer a promising approach to addressing these challenges, but they also depend on strong base models that face the same aforementioned limitations. To overcome these constraints, we propose AnchorInv following the straightforward and efficient buffer-replay strategy. Instead of selecting and storing raw data, AnchorInv generates synthetic samples guided by anchor points in the feature space. This approach protects privacy and regularizes the model for adaptation. When evaluated on three public physiological time series datasets, AnchorInv exhibits efficient knowledge forgetting prevention and improved adaptation to novel classes, surpassing state-of-the-art baselines.

Figures (8)

• Figure 1: Overview of AnchorInv. In the base session, the training dataset is projected to the feature space, and anchor points for each class are identified and saved in the anchor set memory. In incremental sessions, the anchor set guides the model inversion process to generate representative samples of previously seen classes. The inverted samples and the few-shot training set are subsequently used to finetune the backbone. Anchor points for the new classes from the session are also appended to the anchor set memory, for access in later incremental sessions.
• Figure 2: Overview of FSCIL Evaluation Procedure a) illustrates the FSCIL evaluation process in the existing literature. b) illustrates the FSCIL evaluation process adopted in this work. In Session 1, we create $M$ copies of the base session model. In each incremental session, each copy is adapted with a randomly sampled training set. Each copy is evaluated on the test set and the mean and standard deviation are reported across all $M$ trials.
• Figure 3: t-SNE Visualization of the Anchor Set and the Inverted Samples in Feature Space
• Figure 4: AnchorInv Ablations. We compare the last session's performance on the BCI dataset under varying parameter settings. (a) $M$ Number of anchor points per base class (b) $K$ Number of training samples per incremental class
• Figure 5: Visualization of EEG Samples from NHIE