Learning systems face a fundamental tension: they must generalize across experiences while still discriminating task-relevant details. A new computational model called HOLMES (Hierarchical Online Learning of Multiscale Experience Structure) shows how the brain might solve this problem by building nested categories on the fly, trial by trial.
The research
Ines Aitsahalia and Kiyohito Iigaya, researchers at the California Institute of Technology, published their findings on arXiv (q-bio.NC, March 2026, revised July 2026). They asked whether a single model could combine the incremental updating of online latent-cause models with the multilevel structure of hierarchical Bayesian models.
HOLMES merges a variation of the nested Chinese Restaurant Process prior with sequential Monte Carlo inference. This lets the model perform tractable trial-by-trial inference over hierarchical latent representations — without any explicit supervision about the latent structure. In other words, it discovers categories and subcategories on its own as data streams in.
In simulations, HOLMES matched the predictive performance of flat models while learning more compact representations. It then demonstrated one-shot backward transfer: after learning a new high-level category, it could immediately recognize instances belonging to that category without additional training. In a forward transfer task, HOLMES achieved above-chance outcome prediction for stimuli with never-before-seen feature combinations. It did this by exploiting abstract shape-level representations learned across diverse training instances.
The paper includes four main figures and five supplementary figures. The authors report no human sample size because the work is computational, but the simulations provide proof-of-concept evidence for the model's capabilities.
Why it matters
Your brain constantly faces the same challenge HOLMES addresses: you encounter a new object, and you must decide whether it belongs to a known category or deserves a new one. Traditional models either update slowly and miss hierarchical nuance, or they require offline processing that cannot keep pace with real-time experience.
HOLMES suggests that hierarchical structure can be learned online, without a teacher. That has implications for understanding how humans form abstract concepts — like recognizing that a sparrow and an eagle both belong to the broader category of birds, even if you have never seen an eagle before. It also points toward more flexible artificial intelligence systems that adapt to new data without retraining from scratch.
What you can do
To exercise your own hierarchical learning, try categorizing new information as you encounter it. When you learn a new fact, ask: what broader category does this belong to? What distinguishes it from other members? This simple habit strengthens the same kind of flexible, multilevel representation that HOLMES models. You can also practice one-shot transfer by deliberately recalling a high-level concept after learning a single new example — for instance, after reading about a new animal, immediately think of other animals in its class.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.