The brain learns abstract representations of high-dimensional sensory input, but the plasticity rules that enable such learning are unknown. A new study on the Random Hierarchy Model (RHM) reveals that self-supervised local learning rules can match the data efficiency of full backpropagation—without needing a symmetric error network.
The Research
Researchers from EPFL, led by Ariane Delrocq, Wu S. Zihan, Guillaume Bellec, and Wulfram Gerstner, tested biologically plausible algorithms on the Random Hierarchy Model (RHM)—an artificial dataset that mimics the hierarchical structure of natural data. They compared two families of local learning rules: the first uses direct feedback signals to approximate error propagation from the output layer; the second uses layerwise self-supervised objectives (contrastive and non-contrastive).
Results showed that all rules of the first type failed on RHM tasks due to missing input-specific nonlinearities ('masking') found in backpropagation. However, algorithms of the second type successfully learned the hidden hierarchical structure, achieving data efficiency comparable to supervised backpropagation. Notably, these rules are compatible with known synaptic plasticity mechanisms in the brain.
Why It Matters
This study suggests that the brain may use local, self-supervised learning rules to extract complex hierarchical features from sensory input without relying on global error signals. For anyone curious about their own cognition, this means that your brain likely builds abstract representations efficiently using local plasticity rules—which could inform training methods that leverage self-supervised learning principles.
What You Can Do
You can apply this insight by engaging in tasks that require hierarchical pattern recognition, such as learning a new language, playing strategy games, or practicing a musical instrument—activities that drive self-supervised, local learning in your brain.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
