Researchers at the HUN-REN Wigner Research Centre led by Orbán Gergő, together with Argentine collaborators, have modified variational autoencoders (VAEs) to produce more reliable uncertainty estimates for image-based inferences. The new model, called the Explaining-Away Variational Autoencoder (EA-VAE), helps distinguish whether an observed image change originates from the depicted object or from global image factors such as brightness or contrast.
Why this matters
Accurate uncertainty estimation is crucial in applications where an incorrectly confident decision can have serious consequences. From medical image analysis to autonomous perception, systems that can flag when an input is unusual, noisy, or out-of-distribution enable appropriate human oversight and reduce the risk of harmful errors.
Method and main findings
The study focused on variational autoencoders, a class of deep generative models that encode essential image features into a compact internal representation and can reconstruct images from that representation. The researchers observed that conventional VAE uncertainty estimates can fail under common image degradations — for example, when contrast changes, images are noisy or corrupted, or inputs differ significantly from training examples.
To address these shortcomings, the Hungarian–Argentine team augmented the VAE with a component inspired by classical computer vision approaches. The resulting Explaining-Away Variational Autoencoder (EA-VAE) includes a factor that helps the model separate object-related variations from global image-level changes (such as brightness or contrast shifts). Across multiple image datasets and corruption scenarios, the EA-VAE yielded more consistent and dependable uncertainty estimates than the standard VAE.
Connection to biological vision
An intriguing aspect of the EA-VAE is the emergence of a divisive normalization-like computation within its operation. Divisive normalization regulates a unit’s activity relative to neighboring units’ activities, a principle widely observed in the brain’s visual system. Thus, the development not only improves AI performance but also aligns with a computational mechanism present in biological vision.
Practical implications
Improved uncertainty calibration helps systems indicate when human expert review is needed, for instance in medical diagnostics, industrial inspection, or autonomous sensing under changing lighting, noise, or previously unseen inputs. The approach benefits any image-processing system required to cope with variable acquisition conditions or corrupted data.
Publication and context
The results were published in the IEEE Transactions on Pattern Analysis and Machine Intelligence. The work contributes to NeuroAI’s dual goal of using insights about the brain to build better artificial systems and using artificial systems to suggest hypotheses for neuroscience. The project was carried out at the HUN-REN Wigner Research Centre in collaboration with Argentine partners, led by Orbán Gergő.
