Cognition relies on internal representations of relevant information that are organized by constraining population dynamics to activity subspaces referred to as neural manifolds. Here, to examine how manifold geometry is modified by experience, we trained juvenile and adult zebrafish in an odor discrimination task and measured population activity in telencephalic area pDp, the homolog of piriform cortex. No obvious signatures of attractor dynamics were detected; however, olfactory discrimination training selectively enhanced the separation of neural manifolds representing task-relevant odors from other representations, consistent with predictions of autoassociative network models endowed with precise synaptic balance.
Analytical approaches using the framework of manifold capacity revealed multiple geometrical modifications of representational manifolds that supported the classification of task-relevant sensory information. Manifold capacity predicted odor discrimination across individuals, indicating that representational geometry is linked to behavior. Hence, pDp and possibly related recurrent networks store information in the geometry of neural manifolds, resulting in joint sensory and semantic maps that may support distributed learning processes.
Learning generates organized representations of relevant information in the brain that generalize to novel inputs and serve as a basis for cognition. Representational learning is thought to modify synaptic connectivity in memory networks that map patterns of input activity to specific subspaces of neuronal state space according to semantic relationships between inputs1,2. One way to generate such mappings is through attractor dynamics.
Classical models predict that learning enhances recurrent excitation among specific ensembles of neurons, resulting in convergent attractor dynamics that supports the classification of information by pattern completion3,4,5,6,7. Moreover, recurrent networks can give rise to a continuum of stable attractor states as observed, for example, in cognitive maps of space4,8. Alternatively, representational learning may be mediated by mechanisms that do not rely on attractor dynamics.
For example, learning may modify the geometry of neuronal state space and, thus, define neural manifolds that represent relevant information without being attractor states1,2,9,10,11,12,13,14,15; however, as geometrical analyses of neuronal population activity remain challenging, the organization and experience-dependent plasticity of representational neural manifolds in memory networks are still poorly understood. Representational memory has been proposed to be a primary function of piriform cortex, a paleocortical brain area that receives distributed sensory input from the olfactory bulb. Within piriform cortex, excitatory neurons are recurrently connected through an ‘association fiber system’ that is plastic and under neuromodulatory control16,17,18,19.
These observations gave rise to the hypothesis that learning results in the formation of neuronal assemblies that represent odor objects and mediate pattern classification by convergent dynamics20; however, recent analyses of neuronal activity in piriform cortex did not reveal obvious signatures of attractor states. For example, odor-evoked activity is not persistent but phasic and curtailed by inhibition21,22,23. Moreover, while convergent dynamics are expected to reduce variability, intra- and inter-trial variability of neuronal activity seems high in comparison to inputs from the olfactory bulb23,24,25,26.
Extract — continue reading at the source.
