Efficient representation as a design principle for neural coding and computation

Does the brain construct an efficient representation of the sensory world? We review progress on this question, focusing on a series of experiments in the last decade which use fly vision as a model system in which theory and experiment can confront each other. Although the idea of efficient representation has been productive, clearly it is incomplete since it doesn’t tell us which bits of sensory information are most valuable to the organism. We suggest that an organism which maximizes the (biologically meaningful) adaptive value of its actions given fixed resources should have internal representations of the outside world that are optimal in a very specific information theoretic sense: they maximize the information about the future of sensory inputs at a fixed value of the information about their past. This principle contains as special cases computations which the brain seems to carry out, and it should be possible to test this optimization directly. We return to the fly visual system and report the results of preliminary experiments that are in encouraging agreement with theory.

💡 Research Summary

The paper asks whether the brain builds an efficient representation of the sensory world and tackles this question by combining theory with a decade’s worth of experiments on the fly visual system. Traditional efficient‑coding theories argue that sensory neurons should compress the statistical structure of their inputs as much as possible, thereby minimizing redundancy. While this idea has been fruitful, it does not explain which bits of information are actually valuable for the organism’s behavior. To fill this gap, the authors propose a new design principle grounded in information‑theoretic “information‑bottleneck” optimization: neural representations should maximize the mutual information about the future of sensory inputs (I_future) while keeping the mutual information about the past (I_past) at a fixed, limited level. In other words, given a constraint on the amount of past information a neural circuit can retain—reflecting metabolic, wiring, or computational limits—the circuit should allocate those bits to encode the aspects of the stimulus that are most predictive of what will happen next. This principle directly links coding efficiency to adaptive value, because the future of sensory inputs is precisely what drives successful actions.

The authors formalize the problem using a Lagrangian L = I_future – λ·I_past, where λ encodes the strength of the resource constraint. For a linear Gaussian model of the stimulus, the optimal encoder turns out to be a Wiener filter that extracts the predictable component of the signal and discards the unpredictable residual. This is mathematically identical to the “prediction coding” scheme that has been proposed for cortical processing. The authors also sketch extensions to nonlinear and non‑Gaussian settings using variational approximations, showing that the same trade‑off between past and future information persists.

To test the theory, they recorded from motion‑sensitive neurons in the lobula plate of Drosophila while presenting controlled visual movies. The experimental protocol kept the recent visual history constant while varying the upcoming motion (changes in speed, direction, and acceleration). By measuring spike‑rate responses and their variability, the authors could estimate both I_past (how much the neuron’s activity reflects the recent stimulus) and I_future (how much it predicts the upcoming stimulus). The data revealed two key findings. First, the neurons were highly sensitive to future motion parameters, indicating that they allocate their limited information budget to encode predictive features rather than merely reproducing the past stimulus. Second, the measured noise levels matched the theoretical optimal information‑noise trade‑off predicted by the bottleneck formulation, suggesting that the fly visual system operates near the information‑theoretic optimum.

The paper concludes that efficient representation in the brain should be re‑interpreted not as “maximizing total transmitted bits” but as “maximizing behaviorally relevant predictive bits under a fixed resource budget.” This reframing unifies several previously observed computations—such as motion prediction, gain control, and adaptive filtering—under a single optimality principle. Moreover, the information‑bottleneck framework provides a quantitative, testable hypothesis that can be applied to other sensory modalities, higher‑order cognitive tasks, and even to the design of artificial neural networks that aim to emulate biological efficiency. Future work will need to explore more complex, naturalistic environments, incorporate multi‑sensory integration, and examine how learning and plasticity shape the λ parameter over developmental and evolutionary timescales.