Periodic Reporting for period 3 - VarPL (Specificity or generalization? Neural mechanisms for perceptual learning with variability)
Reporting period: 2022-02-01 to 2023-07-31
Our visual system is equipped with a powerful plasticity mechanism, perceptual learning, which serves to improve perception of consistent inputs. However, the signals the visual system receives are extremely variable. How variability affects perceptual learning is unclear. In VarPL, we ask how the visual system tackles the challenge of variability for learning: variability could impair perceptual learning, or, like in language and motor learning, result in the ability to generalize from trained to new materials. To create effective training programs, e.g. for clinical applications, it is crucial to know how to reap the benefits of variability, or, conversely, to overcome the challenges variability poses. Yet, the neural mechanisms by which the visual system copes with variability are unknown, hampering this endeavor. To close this gap, we pursue a new theory, derived from the architecture of cortex: we hypothesize that perceptual learning is not limited to early visual areas, as traditionally envisaged. Instead, learning flexibly occurs at a ‘sweet spot’ along the visual hierarchy whose functional properties match the variability in the given environment. To test this theory, we investigate the role of stimulus variability during training. We hypothesize that variability during multi-day training pushes plasticity to higher brain areas in the visual processing hierarchy, whose tuning properties support generalization. While previous studies in visual learning have shied away from variability, we actively explore its role, and unravel previously unknown flexibility of the visual system that provides a path forward to application of perceptual learning in applied settings, e.g. for vision restoration.
To test our hypothesis, we have developed novel training paradigms to prompt generalizable learning. With these new paradigms, we have been able to determine under which conditions learning is specific, and when it generalizes beyond the training material. For example, we have characterized effects of stimulus variability during training on generalization along several visual feature dimensions. Furthermore, by comparing human behavior to computer simulations, we could provide evidence towards the implementational hypothesis of VarPL, namely that invariant neurons form the basis of these learning and generalization effects. We have also determined whether beyond stimulus features, learning is also specific for effectors, e.g. the kind of movement with which the task is executed. With the use of functional Magnetic Resonance Imaging (fMRI), we currently map out the brain areas that underlie these behavioral effects. Going forward, we plan to use electrophysiological recordings and perturbation techniques to determine learning effects directly on the neural level and assess their causal relevance for behavior.
If successful, VarPL will establish variability as a fundamental organizing principle of learning, enabling a reconceptualization of perceptual learning as generalizable. It will provide insight into the phylogeny of one of our most important visual skills and direct translation of mechanistic insight between primate species fMRI. This way, we hope to enable the development of rehabilitation programs for visual impairments like macular degeneration or after brain lesions. Finally, understanding generalization may further the development of learning algorithms, which still lack our flexibility to generalize.