Optimal predictive coding in a mammalian retina – OPC

A major challenge in sensory neuroscience is to understand the computations performed by different neurons within a given sensory area. For example, in the retina, previous work has shown that there are more than 30 types of retinal ganglion cell (RGC), each of which respond to a very specific aspect of visual scenes. Understanding what is computed by each type is a crucial challenge.  
         A central theory that has been used to understand sensory function is the ‘efficient coding hypothesis’. According to this hypothesis, sensory circuits (e.g. the retina or low-level visual cortex) have evolved to maximise the information transmitted to the rest of the brain, despite internal constraints (such as internal noise and/or consumed energy). This theory has been successful in predicting general features of retinal organisation, such as the shape of receptive fields, their overlap and ON/OFF cell-type ratio.
         However, many major features of RGC responses are not explained by the efficient coding theory. First, the theory cannot account for the large diversity of RGC types. Second, it cannot account for recent results, showing that retinal responses exhibit diverse forms of adaptation to the temporal dynamics of visual stimuli. For example, when presented with a periodic stimulus, RGCs respond to pattern violations, in many cases firing spikes when a pattern is altered or stopped. Moreover, the observed form of adaptation varies greatly between neurons: while some neurons respond strongly to ‘surprising’ stimuli, others show markedly different types of adaptation, that cannot be explained by efficient coding.
         To address these issues, we will develop a new information-theoretic framework of optimal predictive coding, and conduct new experiments to test it in the retina. The proposed theory is a generalisation of efficient coding (which it includes as a special case), and is based on the idea that neural systems (and the retina in particular) should not encode all information equally, but prioritise information that is predictive about the near future. In this theory, neural responses are optimised to maximise information that is predictive about the stimulus at some point in the future (the ‘prediction horizon’), constrained on the information encoded about previous stimuli (the ‘coding capacity’). An important aspect of the theory is that it contains multiple degrees of freedom (e.g. the ‘prediction horizon’ & ‘coding capacity’), that together describe the functional goals and constraints faced by each cell. These extra degrees of flexibility promise an enormous advantage in capturing the diverse set of adaptive behaviours observed in the retina: by varying different factors of our model, we can obtain a simple yet flexible description of different forms of retinal adaption, which emerge as a consequence of different cell types performing different functional goals.
       We will use our theory to investigate RGC responses to complex temporal sequences of illumination. Specifically, we will investigate whether the diverse types of neural adaptation to such stimuli observed previously can be explained by hypothesising that different cell types perform different functional roles and constraints. To test this, we will conduct new experimental recordings of RGC responses to complex temporal sequences, and use our framework to infer the functional objective/constraints that best account for the responses of each cell. Finally, we will use the resulting model fits to predict neural responses to novel stimuli, to which they have not yet been exposed.
         Our approach promises a new way of understanding the diversity of neural responses in a single sensory area, in terms of the diversity of computations performed by different cell types. Such an understanding would go beyond ‘descriptive’ models of retinal responses, and allow us to understand, functionally, what different neurons within the retina are ‘doing’.