Learning from almost nothing: How neural networks survive heavy input corruption
Researchers have explored how neural networks can maintain accuracy even when presented with heavily corrupted input data. Their experiments with multi-layer perceptrons showed that networks could still perform well above chance when over 90% of the input was corrupted, far beyond human recognition. An analysis of infinite-width networks revealed that these networks implement a universal prototype rule, essentially assigning inputs to the class whose training-set average they most closely resemble, explaining this surprising robustness. AI
IMPACT Demonstrates a fundamental capability of neural networks to generalize from noisy data, potentially impacting how models are trained and deployed in real-world scenarios with imperfect data.