A new study from researchers at the Weizmann Institute of Science reveals a simple structural feature that dramatically boosts the computational power of neural networks: local short cycles. Published on arXiv in June 2026, the work by Tom Talpir and Elad Schneidman shows that adding small loops of connections (2- or 3-node cycles) transforms mediocre networks into powerful computing machines.
What the researchers did
The team systematically trained thousands of recurrent neural networks—artificial systems inspired by brain circuits—on a wide variety of Boolean logic functions. For small networks, they built complete “catalogs” of performance, testing every possible wiring pattern. They found that most architectures performed poorly, but those containing local 2- and 3-cycles were far more capable. These cycles often appeared in the minimal network that could solve a given function. The effect was so reliable that a small set of structural statistics—including cycle count—could accurately predict a network's performance.
Moving to larger networks, the pattern held. Typical large networks failed even to approximate randomly chosen functions. However, adding just a few sparsely connected, biologically inspired interneurons—modeled after inhibitory cells in the brain—dramatically increased capacity. These interneurons, when wired to create short cycles, outperformed acyclic or reachability-matched control networks. The authors conclude that local cycles are a general design principle linking neural connectivity to computational power.
Why it matters for your brain
Your brain is a recurrent neural network of about 86 billion neurons. This research suggests that the presence of local feedback loops—where a neuron's output quickly circles back to influence itself or its neighbors—may be a key factor in cognitive performance. While the study used artificial networks, the principles are directly inspired by biology. Understanding that short cycles enhance computation could explain why some brains are more efficient at complex reasoning, memory, or learning. It also hints that brain-training exercises that strengthen such loops might improve cognitive abilities.
What you can do
To support your brain's natural computational abilities, engage in activities that require rapid feedback and pattern recognition. Games like chess, memory puzzles, and learning a musical instrument involve tight loops of perception and action, potentially strengthening local cycles in your neural circuits. Regular practice with varied cognitive tasks may help maintain or enhance these structural advantages.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
