Deep learning models trained without any labels still spontaneously organize data into class-like clusters, according to new research from Patrick Krauss and colleagues at Friedrich-Alexander-Universität Erlangen-Nürnberg. The team's findings, posted on the arXiv preprint server in August 2026, show that as information passes through successive layers of a deep belief network (DBN), the internal representations become increasingly separated according to the true categories of the data.
The Research
Krauss and his team trained deep belief networks on three popular image datasets: MNIST (handwritten digits), Fashion-MNIST (clothing items), and KMNIST (Japanese characters). These networks learn to reconstruct their inputs layer by layer, but crucially, they never see any class labels. After training, the researchers measured how well the networks' internal representations separated the known classes using a metric called the Generalized Discrimination Value (GDV), along with supervised probes and other analyses.
Strikingly, class-specific clustering consistently increased with network depth across all datasets and network widths. The effect was not a trivial artifact: control experiments ruled out random transformations, weight distributions, dimensionality reduction, or neuron saturation as explanations. The first hidden layers often made class identity more accessible to linear and nonlinear probes, while deeper layers produced more compact, prototype-like representations as neurons aligned their feature directions.
The study reveals complementary insights: while average clustering improved with depth, probe accuracy sometimes dipped for a few difficult class pairs, indicating that the network's internal organization does not perfectly match all class distinctions.
Why It Matters
This research offers a fascinating glimpse into how brains—and machines—might organize knowledge without explicit instruction. It suggests that even in the absence of labels, structure inherent in the input can be amplified through hierarchical processing. For human cognition, this echoes how we naturally categorize objects and concepts through experience, not just formal teaching. Understanding this emergent order could inform AI designs that better mimic human learning, and might even offer insights into how our own brains form concepts from raw sensory input.
What You Can Do
While you can't control the hidden layers of your brain, you can encourage your mind to build richer and more flexible categories. Engage in diverse learning: expose yourself to new fields, practice distinguishing subtle differences (e.g., wine tasting, birdwatching, or playing an instrument). The more varied and structured your experiences, the more likely your brain will form robust and accessible concepts.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
