In traditional machine learning, we operate under a "closed-world" assumption: the categories the model sees during training are the only ones it will ever encounter. However, real-world data is non-stationary. A dialogue system might encounter new user intents, or an autonomous vehicle might see a new type of roadside object.
Researchers are moving toward Continual Generalized Category Discovery (C-GCD), a paradigm where models must simultaneously classify known categories and discover entirely new ones from a continuous stream of unlabeled data. This requires balancing the acquisition of new concepts with the preservation of old ones—all without knowing in advance how many new categories exist.
From Static Discovery to Continuous Streams
Early work in Novel Category Discovery (NCD) focused on a two-stage process: learning from a labeled set of "base" classes and then clustering a static unlabeled set to find "novel" classes. Kai Han and colleagues in Automatically Discovering and Learning New Visual Categories with Ranking Statistics demonstrated that models could use ranking statistics to generate pseudo-labels for these new categories, allowing for joint optimization of known and unknown classes.
But static batches are rare in the wild. In a streaming environment, data arrives in sequential stages. Xinwei Zhang and others proposed the Grow and Merge (GM) framework in Grow and Merge: A Unified Framework for Continuous Categories Discovery to handle this. Their approach alternates between a "growing" phase, where the model expands its architecture to accommodate new clusters, and a "merging" phase, which consolidates redundant categories to keep the model efficient.
Preventing Category Collapse and Forgetting
The primary technical hurdle in incremental discovery is "catastrophic forgetting." As a model learns to represent a new category, it often overwrites the features used to recognize previous ones.
To combat this, K J Joseph and Piyush Rai introduced Spacing Loss in Spacing Loss for Discovering Novel Categories. By mathematically forcing different categories to occupy equidistant "anchor points" in the latent space, they prevent the model's representations from collapsing into a single cluster. Similarly, Hyungmin Kim utilized Proxy Anchors in Proxy Anchor-based Unsupervised Learning for Continuous Generalized Category Discovery to provide stable targets for the model to cluster around as new data arrives.
Another strategy involves architectural isolation. In Adaptive Discovering and Merging for Incremental Novel Class Discovery, Guangyao Chen and Peixi Peng decoupled feature learning from category assignment. Their "branch-and-merge" strategy allows the model to spawn new classification branches for novel data while keeping the original branches for known data intact.
Estimating the Unknown
In a true open-world scenario, the model doesn't know how many new categories are hidden in the unlabeled stream. If a model expects three new classes but encounters ten, it will be forced to merge distinct concepts, leading to poor performance.
Bingchen Zhao and Kai Han addressed this in PromptCCD: Learning Gaussian Mixture Prompt Pool for Continual Category Discovery. They used a Gaussian Mixture Prompting (GMP) module that dynamically adapts the model's internal memory. This allows the system to estimate the number of new categories on-the-fly by analyzing the distribution of features in the latent space.
Similarly, Kaidi Cao and Jure Leskovec developed ORCA, which uses an "uncertainty adaptive margin" to prevent the model from being biased toward known classes, as described in Open-World Semi-Supervised Learning. This margin effectively creates "room" in the probability space for the model to say, "I recognize this is something new," rather than forcing it into a known label.
Towards Lifelong Autonomous Learning
The ultimate goal of this research is to create agents that learn like humans. Christoph Käding and Joachim Denzler demonstrated a prototype of this with WALI in Watch, Ask, Learn, and Improve: a lifelong learning cycle for visual recognition. By combining incremental updates with active learning—where the model asks for human help only when it is most confused—they showed that systems could successfully categorize new species in real-time over extended periods.
By moving away from static, closed-world datasets and toward continuous, open-world discovery, these researchers are building the foundation for AI that can evolve alongside the complex, changing environments they inhabit.