Brain-inspired randomness makes AI vision more robust, in lab models

A vision system can read a street sign in clear daylight and fail on the same sign half-hidden by noise. People rarely make that mistake. A team at the Hertie Institute for Clinical Brain Research in Tübingen set out to find out why, and then to test whether the answer could be copied into software. Its findings, published in the journal Patterns, show that two brain-inspired features work better together than apart: the study reports that the combination makes image recognition considerably more reliable.

The first feature is what neuroscientists call top-down feedback. When we see, information does not simply flow from the eyes to the brain. Higher brain areas also send signals back to earlier processing stages, shaping how incoming sensory information is handled. Most artificial neural networks have no such return path. The second feature is neural variability, the random fluctuations in the activity of nerve cells. The team, led by professor Markus Siegel with first author Dr. Antonino Greco, asked whether feedback becomes useful only when combined with that randomness.

The experiment was simple to describe. The researchers trained neural networks to recognize images, some with feedback and some without. To imitate the brain's randomness in a simplified way, they temporarily deactivated individual network units at random during training. Feedback on its own brought little benefit. With the random deactivation added, the networks became substantially more robust to image noise and to targeted changes that cause misclassification, and they also performed better on images that differed from the training data. "We assumed that top-down feedback alone would be sufficient," Greco says. "But that assumption was not borne out. Feedback only becomes effective once the network has learned to cope with randomness."

The researchers' analyses point to a division of labour. Random variability keeps individual units from becoming too dependent on one another during learning, and it makes the network's activity more irregular. Feedback then uses information from higher stages to organize that activity into more stable, compact patterns. The two mechanisms do different jobs, and the study suggests they complement each other.

For engineers who build vision systems, the immediate gain is a training method to test, not a finished product. The researchers present the approach as a possible way to make artificial vision more robust to image disturbances in the future, and the work also serves brain research, since such models help clarify how the brain processes information. The networks in the study were simulations, and the random deactivation is only a simplified stand-in for the brain's variability. The study does not report tests in cars, medical equipment or other real-world devices, and it gives no timeline for that step.