Abstract
The best performing computer vision systems are based on deep neural networks (DNNs). A study in this issue of PLOS Biology shows that DNNs trained on noisy stimuli are better than standard DNNs at mirroring both human behavioral and neural visual responses.
Introduction
Deep Neural Networks (DNNs) currently represent the state-of-the-art in modeling human vision. However, a significant "robustness gap" persists: while a standard DNN can classify clean images with high accuracy, its performance often collapses when faced with minor distortions, such as pixelated noise, that humans easily ignore. This divergence suggests that artificial and biological systems employ fundamentally different strategies for object recognition. A recent study by Jang and colleagues explores whether training DNNs on noisy stimuli can bridge this gap, forcing artificial models to adopt internal representations and behavioral patterns more aligned with the human ventral visual stream.
Methodology
The researchers investigated the impact of noise training on DNN architectures. Models were exposed to two primary types of distortions: spatially uncorrelated Gaussian pixelated noise and spatially correlated Fourier scrambled noise.
To evaluate the "human-likeness" of these models, the study employed three key metrics:
- Recognition Thresholds: Determining the Signal-to-Signal-plus-Noise Ratio (SSNR) at which identification fails for both humans and AI.
- Relevance Maps: Comparing the diagnostic regions used by AI (via attribution heatmaps) with those identified by humans (via "painting" informative regions).
- Neural Alignment: Using fMRI data to correlate DNN layer activity with patterns in the human ventral visual stream—the brain's primary pathway for object recognition.
The models were trained on standard datasets like ImageNet, supplemented with noisy variants, and tested on generalization sets featuring natural weather conditions (e.g., fog, rain) to assess real-world robustness.
Results and Insights
The study found that noise training does more than just improve error rates; it fundamentally shifts the network's internal logic. Noise-trained DNNs achieved recognition thresholds much closer to human performance and, crucially, began to make the same types of errors as human observers.
Furthermore, functional neuroimaging revealed a significant increase in "brain-model correspondence." Noise training forced the networks to abandon "shortcuts"—such as relying on local textures—in favor of the global features and diagnostic regions used by the human brain. This alignment was particularly pronounced in high-level visual areas responsible for complex object recognition.
Discussion
While noise training significantly improves the alignment between artificial and biological vision, it is not a complete solution. There remains a "generalization paradox": while these models handle natural weather better, they still differ from humans in how they generalize to entirely novel distortions. This suggests that the robustness of human vision may stem not only from diverse training data but also from inherent architectural "inductive biases" that current DNNs lack. Future research must determine whether further improvements require more sophisticated training regimes or fundamental changes to network architecture to truly replicate the resilience of the human visual system.
Summary updated