Patterns of neuronal activity underlying behavioral decisions

National Institute Of Mental Health

Investigators

Linked publications, trials & patents

Abstract

Cortical neurons fire in complex patterns of activity during behavior and cognition. In sensory regions of the cortex, animals' interactions with the sensory world evoke neural activity in the cortex. But not all patterns of cortical activity can be elicited by sensory stimuli. We study how non-sensory, artificially induced activity patterns can be used by animals to make behavioral responses. These studies shed light on the limits of cortical function, and how circuit properties like neural connections constrain the set of activity patterns the cerebral cortex can process. In this project period, we have made three significant accomplishments. First, we have continued to develop cutting-edge approaches to perturb patterns of neural activity in the brain. These patterned stimulation approaches allow us to change the activity of groups of neurons — thought to be the fundamental computational unit in the brain — and study effects on network responses and on behavior. Recent innovations involve using mesoscopic imaging methods to expand the number of neurons we can access. Second, we have used these tools to change activity of individual neurons in working brains and measure neurons’ activation functions, the function that describes how they respond to input. Activation functions are key elements of brain neurons and also the neuron-like units of AI systems. Third, we have discovered a fundamental computation, or transformation, performed by cerebral cortical networks. That computation, active filtering, or sequence filtering, allows the brain to process the time dimension of the world. We find that the cortical network learns the structure of the natural sensory world, encodes that structure in local recurrent synaptic connections (weights), and via those connections amplifies sequences of input that correspond to natural vision. This is a form of predictive processing, or predictive coding, that explains how sensory cortex works during natural sensation. First, we have developed tools to enable the control of activity in the brain with single-cell resolution. State-of-the-art all-optical systems promise unprecedented access to neural activity in vivo, using multiphoton optogenetics to allow simultaneous imaging and control of activity in selected neurons at cellular resolution. However, to achieve wide use of all-optical stimulation and imaging, simple strategies are needed to robustly and stably express opsins and indicators in the same cells. Here, we describe a bicistronic adeno-associated virus (AAV) that expresses both the fast and bright calcium indicator jGCaMP8s, and a soma-targeted (st) and two-photon-activatable opsin, ChrimsonR. With this method, stChrimsonR stimulation with two-photon holography in the visual cortex of mice drives robust spiking in targeted cells, and neural responses to visual sensory stimuli and spontaneous activity are strong and stable. Cells expressing this bicistronic construct show responses to both photostimulation and visual stimulation that are similar to responses measured from cells expressing the same opsin and indicator via separate viruses. This approach is a simple and robust way to prepare neurons in vivo for two-photon holography and imaging. Second, we have measured the activation functions of cortical neurons and found them to be similar to the activation functions used in some AI systems. The relationship between neurons’ input and spiking output, the activation function or input-output function, is central to brain computation. Studies in vitro and in anesthetized animals suggest that nonlinearities emerge in cells’ input–output (IO; activation) functions as network activity increases, yet how neurons transform inputs in vivo has been unclear. Here, we characterize cortical principal neurons’ activation functions in awake mice using two-photon optogenetics. We deliver fixed inputs at the soma while neurons’ activity varies with sensory stimuli. We find that responses to fixed optogenetic input are nearly unchanged as neurons are excited, reflecting a linear response regime above neurons’ resting point. In contrast, responses are dramatically attenuated by suppression. This attenuation is a powerful means to filter inputs arriving to suppressed cells, privileging other inputs arriving to excited neurons. These results have two major implications. First, somatic neural activation functions in vivo accord with the activation functions used in recent machine learning systems. Second, neurons’ IO functions can filter sensory inputs—not only do sensory stimuli change neurons’ spiking outputs, but these changes also affect responses to input, attenuating responses to some inputs while leaving others unchanged. Third, we have developed a hypothesis that the extensive recurrent connections within cortical areas filter and sequences of input in sensory cortex, amplifying those that correspond to natural vision. In other words, the cortical recurrent network learns the natural statistics of the world. This learning occurs through experience or development or a combination of the two. Our data shows the recurrent cortical network performs predictive processing, boosting activity patterns associated with the natural world. In daily life, organisms interact with a sensory world that dynamically changes from moment to moment. Recurrent neural networks can generate dynamics, but in sensory cortex it has been unclear if any dynamic processing is supported by the dense recurrent excitatory-excitatory network. Here we show a new role for recurrent connections in mouse visual cortex: they support powerful dynamical computations, but by filtering sequences of input instead of generating sequences. Using two-photon optogenetics, we measure neural responses to natural images and play them back, finding inputs are amplified when played back during the correct movie dynamic context— when the preceding sequence corresponds to natural vision. This sequence selectivity depends on a network mechanism: earlier input patterns produce responses in other local neurons, which interact with later input patterns. We confirm this mechanism by designing sequences of inputs that are amplified or suppressed by the network. These data suggest recurrent cortical connections perform predictive processing, encoding the statistics of the natural world in input-output transformations. This predictive processing enables the cortex to be sensitive to the dynamic and changing aspects of the world. This ‘active filtering’ process may underly a variety of brain processes, from understanding visual movies we see to processing speech that we hear. Our results suggest that sequential processing, a major role of recurrent connectivity in artificial intelligence or machine learning systems, may also be a major role of recurrent connectivity in biological brains.