How does learning impact neural networks in the primary visual cortex?

Funding Round: 3 2015-2017

Research Question: How does learning impact neural networks in the primary visual cortex?

Interdisciplinary Approach: This project will investigate learning effects at the neural network level by combining two-photon calcium imaging in animals learning an orientation discrimination task with a state-space analysis approach.

Potential Implications of Research: Our proposal aims to identify a fundamental learning mechanism that leverages the power of large networks. Our results will help to define the scale at which learning effects need to be studied in the cortex.

An old dog can be taught new tricks – or, in the case of the brain, there is evidence for plasticity (experience-dependent changes of cortical function and circuitry) even in adults. One sign of this plasticity is a phenomenon called ‘perceptual learning’ (PL). In PL, an observer is trained to perform a difficult discrimination task, such as deciding whether a line is slightly tilted relative to vertical. Consistently, PL paradigms show increases in performance with training. These behavioral benefits are retained for long periods of time even without additional training, and are thought to be the result of plastic changes in the brain. Because of the long lasting improvements, PL has been discussed as a potential therapeutic tool for improving vision. Indeed, PL paradigms have been shown to improve vision in amblyopic observers, in which vision in one eye is reduced.

Despite repeated demonstrations of the behavioral gains induced by PL, the underlying neural changes and specific brain areas involved are currently debated. Many PL tasks involve very basic stimuli of the kind represented in primary visual cortex (V1), the first cortical stage of visual processing. The behavioral benefits also have signatures indicative of changes in V1, such as a restriction of training gains to trained parts of the visual field, a trained orientation, or a trained eye. All of these parameters are most strongly represented in V1, rather than higher order visual areas. Yet, recordings of neural responses from V1 come to conflicting conclusions: Studies using magnetic resonance imaging or electroencephalography have reported PL induced changes in V1. Both techniques record the aggregate response of large numbers of neurons in a particular brain region. Surprisingly, recordings from individual V1 neurons have found little or no evidence of PL effects.

We propose that the discrepancy in results from single neurons and large numbers of neurons reflect the basic nature of the PL induced changes. We hypothesize that PL is reflected in small changes distributed across a network of neurons, rather than large changes in a small number of neurons. This kind of distributed learning would afford the brain much larger flexibility to adapt, as even subtle changes in individual neurons could amount to large changes at the network level. Measuring these kinds of distributed changes requires suitable measurement and analysis tools. Here, we will use a combination of two state-of-the-art approaches, two-photon calcium imaging and state space analysis. In our experiments, we will train ferrets on a visual discrimination task to induce PL. Ferret V1 shares organizational principles with V1 of humans and monkeys. Using two-photon imaging (a high resolution microscopy technique that can measure responses of many neurons simultaneously), we will follow the same group of neighboring V1 neurons while the animal learns. Whether and how learning influences these neurons will then be determined using state-space analysis, an approach developed to quantify the behavior of a network of neurons.

Our results would demonstrate an important mechanism that allows the brain to learn. The greater flexibility and robustness of network changes, compared to changes in individual neurons, would increase total capacity for learning. Our findings would define the scale at which learning occurs, informing future experiments and more generally models of learning. In addition, our study would provide a direct test of plasticity of local circuits in the adult brain. Lastly, our findings may help with further developing PL as a therapeutic tool.