From the noisy information bombarding our senses, our brains must construct percepts that are veridical – reflecting the true state of the world – and informative – conveying what we did not already know. Influential theories suggest that both challenges are met through mechanisms that use expectations about the likely state of the world to shape perception. Predictive coding, a theoretical framework in which the brain compares a generative model to incoming sensory signals, seeks to explain this inferential process of perception. Implementation of predictive coding at a mechanistic level in individual neurons is however not well understood. This dissertation builds on the field of computational modeling of sensory systems to further our understanding of perceptual inference in individual neurons in the auditory system of songbirds (European starling). Chapter 1 summarizes some possible ways in which predictive coding could be realized at the algorithmic level. The main contribution of this chapter is not to present models that can uniquely account for sensory data, but to reconcile different cognitive behaviors under the umbrella of predictive coding which were previously considered incompatible. Chapter 2 combines computational modeling with electrophysiological activity to study internal representations of sensory information under predictive coding framework in single auditory neurons in songbirds. This study examines unique components of the response variance guided by predictive coding. It is shown that during song listening, individual neuron responses in primary and secondary auditory regions are modeled by expectations of future song and uncertainty of song. Generative machine learning approaches enable us to make hypotheses about internal generative models for complex natural stimuli. Chapter 3 explores the flexibility of the internal model and whether prediction error is correlated with behavior. This study draws primarily upon the complex structure of birdsong, developing methods to analyze the acoustic and temporal structure in vocal signals and then behaviorally and physiologically probing the underpinnings of the predictive structure of birdsong. The findings in this chapter indicate that prediction error encoding is modulated by task-relevant generative models and behavior. Taken together, my dissertation work offers a unified explanation of inferential frameworks of perception in contexts of systems processing behaviorally relevant signals, showcasing its adaptation to these systems.

Characterization of response properties of neurons in higher-level sensory areas is not well defined. Here we show that firing rates of neurons in a secondary sensory forebrain area of songbirds can be modeled by different representations of birdsong. In this work, we modeled neurons in the caudo-medial nidopallium (NCM) of adult European starlings with three different representations of the natural birdsong called signal, prediction, and error. Prediction spectrogram was computed by training the data as a Gaussian distribution on a loss function given by the negative log likelihood, and then estimating the means and variances of the signal. Using our Maximum Noise Entropy (MNE) model, responses were predicted by the logistic function, the parameters of which are obtained from the MNE model. Predictions of neural responses were computed by using both a full MNE model, and then by only considering the linear parameters of the model. The neural responses to natural stimuli obtained using prediction and error MNEs were close to the actual response in the NCM. The concept of stimulus representations obtained from predictive coding models may be useful for modeling neural responses in higher-order sensory areas whose functions have been poorly understood.