UC Berkeley

Sensory information is encoded with sparse spiking in rodent sensory cortex, but the organization and functional basis of this sparse code is not well understood. I conducted two studies to characterize function and learning in the cortex. In the first study, I used population calcium imaging to study the circuit-level factors underlying sparse coding in layer (L) 2 of mouse somatosensory cortex. Whisker deflection elicited low-probability spikes in small, shifting neural ensembles spanning multiple cortical columns. Neurons within a column-sized imaging field were tuned heterogeneously to many different whiskers, contrary to standard models of somatotopy. A spectrum of whisker-evoked response probability existed across neurons that correlated strongly with spontaneous firing rate. This correlation indicates that a major component of responsiveness is independent of experimental stimulus choice. The distribution of responsivity was skewed, indicating the existence of a small population of highly-responsive neurons. Highly-responsive neurons included pyramidal cells and interneurons, and individual whisker deflections were primarily encoded by a small, stable population of highly responsive cells. L2 neurons projecting to motor (M1) and secondary somatosensory (S2) cortex differed in whisker tuning and sparseness, suggesting these intermingled populations send disparate information to their targets. Thus, sparse coding in L2 reflects heterogeneous sensory tuning, low average response probability across neurons, a skewed distribution of inherent responsiveness that includes a small number of more-active neurons, and functional specialization of S1 output streams.

In order to test whether the large pool of unresponsive neurons observed might be important in learning, I developed a novel type of brain-machine interface (BMI) based on calcium imaging in the intact cortex. In this BMI task, the mouse learned to use voluntary modulations of neural activity to control a device. The BMI design allowed for direct control over the relationship between neuronal activity and behavioral output. We trained mice to operantly control an auditory cursor using spike-related calcium signals recorded with 2-photon imaging in motor and somatosensory cortex. Mice rapidly learned to modulate activity in layer 2/3 neurons, evident both across- and within-sessions. Learning was accompanied by striking modifications of firing correlations within spatially localized networks at fine scales (10-100 microns). We found that less-active neurons, and even silent neurons, could dramatically up-modulate their firing to successfully learn the task. Neurons in a `cloud' around the BMI-controlling neurons initially exhibited task-related activity, which dampened out as the animal honed in on the specific cells controlling the device. This suggests an economization of activity, which may be reflective of the sparse firing strategies in the cortex.

These studies both point to the existence of a gradient of activity in cortical neurons in L2/3, which can nevertheless be volitionally manipulated by learning. Neurons in L2/3 of rodent somatosensory cortex had unexpectedly divergent tuning, organized in a salt-and-pepper fashion. The reliability of their responses depended mainly on intrinsic, stimulus-independent responsivity, and more minimally on tuning and downstream targets. However, even extremely inactive neurons could be induced to modify their firing dramatically by coupling their activity to reward. The network rapidly learned to minimize the number of neurons necessary to perform the task, suggesting that an economizing impetus might be at work in superficial cortex.