SPRINT: Semi-supervised Prototypical Representation for Few-Shot Class-Incremental Tabular Learning
| Authors | Umid Suleymanov et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2603.04321 |
| Download | |
| Categories | cs.CV, cs.AI |
Abstract
Real-world systems must continuously adapt to novel concepts from limited data without forgetting previously acquired knowledge. While Few-Shot Class-Incremental Learning (FSCIL) is established in computer vision, its application to tabular domains remains largely unexplored. Unlike images, tabular streams (e.g., logs, sensors) offer abundant unlabeled data, a scarcity of expert annotations and negligible storage costs, features ignored by existing vision-based methods that rely on restrictive buffers. We introduce SPRINT, the first FSCIL framework tailored for tabular distributions. SPRINT introduces a mixed episodic training strategy that leverages confidence-based pseudo-labeling to enrich novel class representations and exploits low storage costs to retain base class history. Extensive evaluation across six diverse benchmarks spanning cybersecurity, healthcare, and ecological domains, demonstrates SPRINT's cross-domain robustness. It achieves a state-of-the-art average accuracy of 77.37% (5-shot), outperforming the strongest incremental baseline by 4.45%.
Engineering Breakdown
Plain English
This paper introduces SPRINT, the first Few-Shot Class-Incremental Learning (FSCIL) framework designed specifically for tabular data streams like logs and sensor readings. FSCIL is a well-established computer vision task where systems learn new classes from very few labeled examples without catastrophically forgetting old knowledge, but nobody has seriously adapted it to tabular domains before. SPRINT combines confidence-based pseudo-labeling to improve representations of new classes with a mixed episodic training strategy, and crucially exploits the fact that tabular data has negligible storage costs compared to images, allowing the system to retain base class history without buffer constraints. The framework addresses a real production gap: tabular streams are everywhere (databases, IoT, event logs) and often come with abundant unlabeled data but scarce expert annotations—exactly opposite the assumptions of vision-based FSCIL methods.
Core Technical Contribution
The core novelty is recognizing and exploiting the structural differences between tabular and image domains for incremental learning: tabular data enables cheap storage, making the restrictive memory buffers of vision-based FSCIL unnecessary. SPRINT introduces a mixed episodic training strategy that leverages confidence-based pseudo-labeling, allowing the model to bootstrap new class representations using unlabeled data—a strategy that works because tabular streams have abundant unlabeled examples but few labels. The authors explicitly depart from vision pipelines that assume storage constraints and instead design around the tabular reality: retain base class history without penalty, use pseudo-labels to enrich novel class embeddings, and train on mixed old-and-new episodes rather than maintaining replay buffers. This is the first framework to properly formalize FSCIL for tabular distributions, treating it as a distinct problem rather than a port of image-based methods.
How It Works
SPRINT operates in phases over incoming tabular data streams where new classes arrive with very few labeled examples. During training on each new task, the system uses a mixed episodic approach: it constructs episodes that blend base class samples (drawn from retained history without storage penalty) and novel class samples (both the scarce labels and pseudo-labeled unlabeled data). The confidence-based pseudo-labeling mechanism works by running the model on unlabeled tabular data, assigning labels only to samples where the model's prediction confidence exceeds a threshold, effectively using high-confidence predictions to create synthetic training data for underrepresented new classes. The model learns to map tabular features into an embedding space where base and novel classes remain separable; because storage is cheap for tabular data, previous base class examples stay available indefinitely, solving the catastrophic forgetting problem through data retention rather than clever regularization. At inference, the system classifies new tabular samples by embedding them and comparing to class prototypes (learned centroids), with class predictions updated incrementally as new examples arrive.
Production Impact
For engineers operating on tabular streams (sensor networks, database transaction logs, user event streams), SPRINT eliminates a painful constraint: you can now continuously learn new patterns without the memory-intensive replay buffers that vision-based FSCIL requires. In a real pipeline, this means your model can automatically adapt when new classes appear in production logs—say, a new error type or a new user segment—using only a handful of labeled examples, while automatically leveraging the ocean of unlabeled data already flowing through your system. The pseudo-labeling mechanism provides a practical safety valve: instead of requiring all new data to be labeled by experts, only high-confidence predictions are used, reducing annotation burden in scenarios where labeling is expensive (domain expert reviews) while maintaining safety through confidence thresholds. Trade-offs include: (1) confidence thresholds must be tuned per domain—too low and you poison training with bad labels, too high and you ignore useful unlabeled data; (2) storage is cheap but not infinite, so retention policies still matter at massive scale; (3) pseudo-labels can accumulate bias if the model makes systematic errors on certain data regions.
Limitations and When Not to Use This
SPRINT assumes that base class data can be retained indefinitely, which breaks down at extreme scale (petabyte-level streams where storage costs matter) or in privacy-critical settings where keeping old data violates regulations. The confidence-based pseudo-labeling strategy assumes the model's confidence calibration is reliable, but many tabular models (especially tree-based or poorly calibrated neural networks) produce overconfident predictions, meaning the threshold heuristic could fail silently in production. The paper focuses on the class-incremental setting where new classes appear sequentially, but doesn't address the harder scenario where class definitions shift gradually or where old and new classes overlap in feature space. The framework also inherits FSCIL's core limitation: it still requires each new class to have at least a few clean labeled examples—truly zero-shot class learning remains out of reach, limiting applicability when even a handful of labels are unavailable.
Research Context
This work extends the Few-Shot Class-Incremental Learning literature, which originated in computer vision and has focused heavily on image benchmarks (miniImageNet, CIFAR-100-FS) where memory constraints force design choices that don't apply to tabular data. SPRINT is positioned as the first framework to properly formalize FSCIL for non-vision domains, opening the door to applying incremental learning to time series, structured logs, and other tabular streams where vision assumptions fail. The confidence-based pseudo-labeling draws inspiration from self-training and semi-supervised learning literature (consistency regularization, entropy minimization) but applies it in the novel context of few-shot class increments where labeled data is both scarce and sequential. By decoupling FSCIL from its vision heritage, the paper suggests a broader research direction: task formulations should be domain-specific, and solutions built for images can inadvertently bake in constraints that handicap other data modalities.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::