Detecting rare but relevant events in systems neuroscience

Abstract

Animals actively move their sensory organs, often in a rhythmic manner, to gather information from the external environment. The movements performed to sense the world are often very subtle and hard to detect in recording devices. For instance, in the visual domain, eye movements with amplitudes smaller than a degree of visual angle can occur. These tiny movements, called microsaccades, are at the threshold of the resolution of most recording techniques and one could be tempted to ignore them when studying vision. Yet, they might play an important role in visual processing. My thesis shows that microsaccades should not be ignored, that an algorithm can detect them accurately, and that the same algorithm can be used to detect any other seemingly “petty” events that deserve to be detected among noisy signals. In the first part, we demonstrated that microsaccades have a long-lasting impact on visual processing. We designed behavioral experiments to probe visual detectability and reaction time for stimuli presented at various moments relative to microsaccade onset. By probing the behavioral performance at multiples time points, we could reconstruct a signal that revealed oscillations occurring during visual processing. These oscillations occurred in the beta and alpha range and were synchronized to microsaccade generation. Moreover, the oscillations were sequential, occurring as two pulses, one in each hemifield, depending on the direction of the microsaccade. We also found that microsaccades are associated with a long-lasting increase in contrast sensitivity for stimuli presented in the same hemifield than their direction. These discoveries were important because they demonstrated that visuomotor processing is almost never exempt from the impact of subtle, seemingly irrelevant, movement behaviors. The results therefore established the need for accurate detection of microsaccades and other potentially significant events in brain activity and behavior. We thus designed, in a second study, a deep neural network that performs human- level eye movements detection even in noisy eye traces. Our algorithm outperformed the state-of-the-art algorithm for eye movement detection as well as many commonly used algorithms. In a third study, we finally showed that our algorithm can be generalized to other types of signals by detecting complex spikes in extracellular recordings of cerebellar Purkinje cells. We demonstrated human-level detection of complex spikes, outperforming commonly used online algorithms. Furthermore, our approach also accurately estimated the duration of complex spikes, which provides important information about the coding of error in the cerebellum. Putting all of the above together, this thesis argues for a careful control of exploratory movements when studying sensory processing. It also provides the tools necessary to approach a problem that is common in many different fields of neuroscience: the detection of an event of interest in a noisy signal.