Skip to main content

Statistical methods for modelling population activity in visual cortex

Final Report Summary - POPCODE (Statistical methods for modelling population activity in visual cortex)

Understanding how neurons in the brain collectively process the sensory input, perform computations and guide behaviour is one of the central goals of neuroscience. A deeper understanding of these processes will have a wide range of scientific and clinical implications, for example for the development of prosthetic devices. For many years, neuroscientists were limited to measuring electrical pulses (called 'spikes') of single neurons, but did not have the tools to assess interactions across neural populations. Recent advances in recording technology now make it possible to monitor the spiking activity of multiple neurons simultaneously, and allow unprecedented insights into the collective dynamics of neural networks. However, understanding the complex data generated by modern recording methods is a challenging task that requires appropriate statistical tools. Multi-cell recordings yield high-dimensional data, which can be hard to visualise and interpret. For example, neural population activity exhibits substantial trial to trial variability with rich, dynamically changing statistical structure.

In this project, we developed novel statistical methods to model population measurements of neural activity, and thereby to facilitate quantitative studies of neural population coding. Of particular interest to us was the question of context in the code, i.e. how the 'meaning' of the activity of a neuron is modulated by the dynamical activity of other neurons. To this end, we designed a statistical method which can capture the effect of underlying cortical dynamics on single-neuron activity. Our results suggest that the observed complex structure of neural activity in cortical populations can be explained using a simple model of cortical population dynamics based on a low-dimensional dynamical system. In particular, we compared our model to a popular statistical model ('coupled generalised linear spike-response models') on multi-neuron recordings from motor cortex. We showed that our latent dynamical approach outperformed this alternative model in terms of goodness-of-fit, and reproduced the temporal correlations in the data more accurately.

More generally, we developed methodology which can deal with the highly dynamical and state-dependent properties of neural activity. Therefore, the tools developed in this project have the potential to be useful in a wide range of different contexts, and will contribute towards unravelling how cortical circuits collectively process information. Finally, our results have implications for the design of decoding algorithms of brain-machine interfaces, as they suggest that appropriate decoding algorithms should also account for the impact of cortical dynamics on neural activity.