Representation and learning in cerebellum-like structures

de Ann Kennedy

Animals use their nervous system to translate signals from their sensory environment into appropriate behavioral responses. In some cases, these responses are hard-wired through genetic sculpting of neural circuits, such that certain stimuli drive innate behavioral responses in the absence of prior experience [Ewert, Burghagen, and Schurg Pfeiffer 1983; Yilmaz and Meister 2013; Wu et al 2014]. But most often, responses to stimuli are modified over the course of an organism's lifetime via associative learning, in which past experience is used to adaptively modify the neural circuits controlling behavior. The remarkable regularity of cerebellar circuitry made it an early target of experiments seeking a link between neural circuit structure and computational function (Eccles, Ito, and Szentgothai, 1967). These efforts led to a first generation of models describing cerebellar cortex as a device for associative learning, remarkable for their focus on linking each cell type of cerebellar cortex to a computational aspect of associative memory formation and adaptive control ([Marr 1969; Albus 1971; Ito 1972). In subsequent decades, specialized neural architecture resembling that of the cerebellum has been identified in several other brain regions, including the dorsal cochlear nucleus of most mammals (Oertel and Young, 2004), the mushroom body of the insect olfactory system (Farris, 2011), and a region evolutionarily and developmentally related to the cerebellum in the brains of weakly electric fish, the electrosensory lobe (Bell, Han, and Sawtell, 2008). This has raised the hope that a similar computational mechanism is at work in these structures. It is not easy to find behavioral paradigms that isolate learning in the cerebellum, and a complete mechanistic account of learning during commonly studied behaviors has remained elusive. In this thesis, I analyze two cerebellum like structures, the electrosensory lobe of the mormyrid fish and the mushroom body of the fly olfactory system, in which mapping out associative learning is more tractable, due to the availability of well controlled learning paradigms and the development of powerful biochemical and genetic techniques. With the help of my experimental collaborators, I constructed computational models of the electrosensory lobe and mushroom body from electrophysiological and anatomical data, and studied the process of associative learning in these models. In both systems, an initial sensory representation is first projected up into a high dimensional space, and then read out via convergent input onto individual neurons. Learning adjusts the input to readout neurons, causing changes in their responses to future stimuli that alters their drive to downstream nuclei. Two details shape how each circuit handles associative learning: the way in which sensory inputs are represented, and the mechanism of learning. Together, these two pieces determine what transformations each circuit is able to learn and how it generalizes after learning. In the four chapters of this thesis I present four related projects dealing with sensory representation and learning in cerebellum-like structures. The first chapter has previously been published as a paper and describes a model for cancellation of self generated sensory input in the passive electrosensory system of the mormyrid fish. In the second chapter, I adapt this model to a more high dimensional cancellation problem in the fish's active electrosensory system, which deals with the effects of the fish's body on the electric fields it generates. In the next two chapters, I construct a network model of odor representation in fly olfactory system, terminating at the mushroom body. Finally, I use this model in conjunction with recent experimental findings on the output of the mushroom body, to build a model of associative odor learning in the fly.