Predictive coding networks capture human neural representations missing in supervised DNNs
Neuroscientific learning theories propose that the brain acquires knowledge by constructing internal world models. Supervised learning, the dominant approach in deep neural networks (DNN), relies on external category labels, making it difficult to reconcile with biological learning. There is an increasing trend towards more biologically valid approaches, such as predictive (minimize future…
Neuroscientific theories suggest that the brain forms internal models to acquire knowledge. In contrast, supervised learning in deep neural networks (DNN) uses external labels, making it hard to align with biological learning processes. Recent research aims for more biologically relevant methods, like predictive or contrastive objectives, but these methods often depend on various DNN architectures, sizes, and hyperparameters, complicating comparisons.
To examine the impact of learning, a study isolated the effects of small, identical networks trained with predictive, contrastive, and supervised objectives, as well as local and global learning. Results indicate that brain representations from statistical learning align better with a predictive local target compared to supervisory or contrastive targets.
During learning, the brain reduces category-specific representations while maintaining predictive ones. Additionally, predictive objectives reveal variance in brain activity that standard supervised DNNs cannot capture, with this variance linked to predictive processing rather than stimulus or mismatch processing. These findings provide a controlled method to explore the algorithmic foundations of learning and highlight prediction as a key learning mechanism.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
