NEURAL MECHANISMS OF SENSORY REPRESENTATIONS AND DECISION MAKING
The research in my lab aims to understand the fundamental principles of neural circuit organization and how this organization relates to the computations that support sensory and behavioural function. The approach we take is to (i) record activity in identified neurons in large ensembles to uncover the computations taking place during sensory processing and sensory-guided behaviours, and (ii) understand how these computations arise from the neural hardware: from the synaptic interactions between identified cell types that differ in the patterns of input and output connectivity.
For this purpose, we focus on sensory processing in visual cortex and connected brain areas of the mouse using a combination of methods, including two-photon calcium imaging in anesthetized and behaving mice, in vitro whole-cell recordings, in vivo whole-cell and extracellular recordings, optogenetics, genetic labelling and anatomical tracing, single-cell transcriptional profiling, visual behavioural tasks, and computational modelling.
1) Principles of connectivity and computation in local networks in visual cortex
Each region of the neocortex comprises a diversity of cell types whose complex interactions govern information processing. The nature of these interactions and resulting computations is poorly understood. Our research has begun to uncover the rules of connectivity between identified cell types that underlie specific computations during sensory processing in visual cortex of the mouse. To this end, my laboratory has developed a method by which visual response properties of neurons are first characterized with two-photon calcium imaging in vivo, and then synaptic connections between a subset of these neurons are assayed with multiple whole-cell recordings in slices of the same tissue (Ko et al 2011 Nature). Using this approach, we discovered several principles of circuit organisation. First, neurons with correlated responses to visual stimuli connect most frequently with strong and reciprocal connections (Cossell et al 2015 Nature). This circuit architecture — comprising strong recurrent excitation within subnetworks of correlated neurons — supports amplification of responses to specific sensory stimuli, and thus promotes robust coding and effective information transmission to multiple postsynaptic targets. Second, the highly structured connectivity between excitatory cells emerges during development via correlation-based learning rules (Ko et al, 2013 Nature). Third, excitatory connectivity in visual cortex is constrained by long-range projection targets, whereby neurons projecting to the same target area are preferentially connected to each other but not to neurons projecting to another target area (Kim et al, unpublished). Thus, neurons belonging to each projection class form mutually exclusive subnetworks that separately process information before relaying it to distinct target areas. Fourth, unlike the specific connectivity between excitatory neurons, a subset of local inhibitory interneurons expressing parvalbumin is interconnected densely and promiscuously to the local network (Hofer et al 2011 Nature Neuroscience). This explains their broad selectivity for sensory features and suggests their main role is normalisation of excitatory neuron responses.
We are extending this work by identifying additional connectivity motifs between different excitatory and inhibitory cell types within and across cortical layers. Our next big challenge is to combine these connectional and functional datasets into a computational model of the visual cortex circuit to create a unified framework for testing the contribution of identified network components to the representation and transformation of visual information (in collaboration with Angus Silver at UCL, and with Maneesh Sahani at the Gatsby Computational Neuroscience Unit).
2) Principles of long-rage connectivity and information transfer between visual cortex and other brain regions
While connections between nearby neurons mediate local information processing within a cortical region, extensive long-range axonal projections are fundamental for exchange of information between brain areas specialised in different functions. Surprisingly little is known about the rules of connectivity and resulting computations in long-range circuits that link cortical areas. This knowledge is crucial for developing conceptual frameworks for understanding what the cortex actually computes, with respect to associations, predictions, and transformations of information in hierarchical networks.
The organisation and function of long-range projections from visual cortex
Information from visual cortex is broadcast via long-range axonal projections to many other cortical and subcortical areas that process different aspects of the visual scene or guide behavioural decisions based on visual input. Virtually nothing is known about the number of unique projection classes and what information they convey to target areas. To address this, we are characterising the diversity of excitatory projection neuron classes - defined by their projection fields in different target areas - using serial-section two-photon microscopy and single-cell whole-brain anatomical reconstructions. This is revealing distinct projection types originating in visual cortex each targeting a unique set of cortical areas. Next, we use imaging in identified projection neurons to determine what information they broadcast to downstream targets during different behavioural tasks. To this end, we have developed several behavioural tasks in a virtual reality environment (Poort et al 2015 Neuron) that engage the visual system in different ways: visual pattern discrimination, visual spatial attention, memory-based visual expectations, motor-based visual expectations and visually-guided action. The hypothesis we wish to test is whether there is dynamic routing of information through selection of most relevant output channels as a function of task demands. Finally, we plan to use optogenetic manipulations to test the requirement of identified projections for perception and behavioural decisions.
The organisation and function of feedback projections to visual cortex
Activity in sensory areas of the neocortex not only reflects the occurrence of external stimuli but also contains information about stimulus context, its behavioural relevance, as well as current goals, expectations, actions, or past experience (reviewed in Harris & Mrsic-Flogel 2013 Nature). These originate in specialised cortical areas, carried to the visual cortex by an extensive network of intracortical long-range projections. Current research in the lab is focused on understanding the circuit mechanisms by which diverse contextual signals are integrated with feed-forward visual information (Poort et al 2015 Neuron), and how such signals emerge through experience. By labelling, imaging and manipulating the candidate long-range projections innervating visual cortex (and related cortical areas) during different visually-guided behaviours, we aim to identify (i) what information is fed back from other cortical areas, (ii) how the nature of this information is modified by task-engagement, attention and expectation, and (iii) the circuit mechanisms by which top-down projections influence visual processing in relation to different excitatory projection types and inhibitory cell classes in visual cortex. To this end, we discovered that long-range feedback projections from a higher visual area preferentially target neurons in primary visual cortex which project to the same target area, thereby forming reciprocal long-range networks (Kim et al, unpublished).
Ultimately this research will provide an insight into the information that is exchanged and the computations taking place between pairs of connected areas as a function of tasks requiring vision to guide behaviour. This will be verified by simultaneously imaging two or more connected areas using large field-of-view two-photon microscopy. This knowledge will be used to constrain models of cortical processing (such as predictive coding in hierarchical networks).