Traditional machine learning models operate under a strict closed-world assumption: all categories the model will ever encounter are predefined during initial training. When a standard classifier encounters something entirely new, it is forced to misclassify it as one of its known categories.
To break this limitation, researchers initially developed Open-Set Recognition (OSR) and Novel Class Discovery (NCD). OSR allows a model to say, "I don't know what this is," by rejecting unfamiliar inputs. NCD goes a step further, taking a static pool of unlabeled data and clustering it to discover entirely new categories.
However, as highlighted by Rizve et al. and Xi et al., these frameworks assume a static environment where all novel data is available simultaneously in a single batch. In the real world, data arrives as a continuous stream. An autonomous agent or an online classifier encounters new categories sequentially over time.
Incremental Class Discovery (also known as Continuous Category Discovery) bridges this gap. It requires a system to simultaneously execute three tasks over an evolving data stream:
- Detect when an incoming instance belongs to an unknown category.
- Discover and cluster multiple distinct new categories from unlabeled data on the fly.
- Integrate these new categories into its existing knowledge base without suffering from catastrophic forgetting of previously learned classes.
The Mechanics of Discovery: Reject, Cluster, and Expand
To understand why Incremental Class Discovery works, we must look at how models dynamically expand their label spaces without human supervision. The process typically follows a pipeline of rejection, representation growth, and consolidation.
1. Rejection and Open Classification
Before a model can discover a new class, it must recognize that an instance does not fit into its current vocabulary. Early foundational work in open text classification by Shu et al. achieved this by replacing the standard softmax layer with a collection of 1-vs-rest sigmoid classifiers. By applying Gaussian-based thresholding to the sigmoid outputs, the model could reject out-of-distribution samples rather than forcing them into a known bin.
2. The Grow and Merge Paradigm
Once unknown samples are isolated, the model cannot simply treat them as a single "unknown" catch-all category. It must discern how many new classes exist within that pool.
Xinwei Zhang formalized this process through the Grow and Merge (GM) framework. When a stream of unlabeled data arrives, the model enters a growing phase, where it instantiates new, tentative category representations to capture underlying data clusters. Because unsupervised clustering can be noisy and generate redundant categories, the model periodically enters a merging phase. During this phase, it computes similarities between the newly formed clusters and existing classes, consolidating overlapping representations to maintain a clean, distinct taxonomy.
[Incoming Data Stream]
│
▼
[Known/Unknown Filter] ──(Known)──► [Update Existing Classes]
│
(Unknown)
▼
[Growing Phase] ──► Instantiates new tentative clusters
│
▼
[Merging Phase] ──► Consolidates overlapping & redundant classes
Managing the Geometry of the Feature Space
The primary reason deep neural networks fail at incremental learning is that updating a model on new data distorts the embedding space of old data. If you force a network to learn a new cluster, the weights shift, and the boundaries of historical classes collapse—a phenomenon known as catastrophic forgetting.
To make incremental discovery work, models must preserve the geometry of old classes while carving out distinct geometric regions for new ones. Researchers achieve this through three main strategies:
1. Dynamic Prompting and Parameter Isolation
Instead of updating the core weights of a large pretrained model, recent approaches isolate historical knowledge by using learnable prompts. In PromptCCD, Zhao et al. introduced a Gaussian Mixture Prompt Pool. Instead of relying on rigid class slots, the framework models categories as mixtures of Gaussians within a prompt space. This allows the system to dynamically adapt its internal memory to represent new classes and estimate the total number of novel categories on-the-fly, without altering the underlying feature extractor.
2. Density-Based Anchoring
Another way to protect the embedding space is to anchor old classes using non-parametric distributions. Zhao et al. demonstrated this in their Incremental Generalized Category Discovery (IGCD) framework. By using density-based sampling, the model identifies core representative prototypes of historical classes. When new unlabeled data streams in, these prototypes act as gravitational anchors in the feature space, ensuring that the contrastive learning objectives pull new data into novel clusters without shifting the coordinates of old ones.
3. Generative Replay
When historical data cannot be saved due to privacy or storage constraints, models can recreate the geometry of the past synthetically. Zhang et al. showed that instead of storing raw images, a model can leverage generative architectures (like Stable Diffusion) to replay basic representations of old classes during new training phases. By focusing the replay on fixing classification boundaries rather than reconstructing complex spatial layouts, the model successfully stabilizes its decision boundaries against representation drift.
Application Across Domains
The necessity of incremental class discovery spans multiple modalities, each requiring unique architectural adaptations to handle streaming data.
- Object Detection: In computer vision, models must localize and classify objects simultaneously. Xiaowei Zhao introduced an Auxiliary Proposal Advisor to help object detectors locate unknown objects, while Wu et al. (UC-OWOD) advanced this by clustering those localized proposals into distinct, classified unknown categories over time.
- Natural Language Processing: Text streams present challenges due to the discrete nature of language. Xia et al. addressed multi-round new class discovery in text classification by framing the problem as textual entailment, allowing models to leverage indirect supervision from large language inference datasets. For dialogue systems, Zhao et al. developed geometry-aware representations specifically optimized to discover novel user intentions from short, conversational text streams.