Decomposing natural stimulus-specific sensory neural information using a validated model of the retina and diffusion models
Steeve Laquitaine*¹, Simone Azeglio¹, Carlo Paris¹, Ulisse Ferrari¹, Matthew Chalk¹
¹ Paris Vision Institute, Sorbonne University, Paris, France
*Email: steeve.laquitaine@inserm.fr IntroductionA key question in sensory neuroscience is how much and what information neurons encode. It is known that neurons in early visual areas respond to stimuli within localized regions of space (receptive fields). Mutual information quantifies the amount of information such neurons encode about all stimuli, but it does not reveal which stimuli contribute most [1,2,3]. The Fisher information measures the sensitivity of such responses to small changes in stimuli [4,5]. However, it does not decompose information in a stimulus-specific manner. Prior decomposition methods exist, but they are hard to interpret, and impractical for high-dimensional stimuli like natural images [3,6,7,8]. Here, we introduce an information decomposition that overcomes these limitations.
MethodsOur method is grounded in information theory and tells us which specific stimuli, or regions of stimuli (e.g. edges, shapes), contribute most to a neuron's response. We require that any valid and interpretable decomposition satisfy three intuitive criteria: (1) it must be a true decomposition of the mutual information (i.e. the total information is the sum of contributions from each stimulus); (2) it must be local (so changes in the response to one stimulus affect only the information assigned to nearby related stimuli) and (3) positive for interpretability. Unlike previous decomposition, our method achieves all three, and it can scale to high-dimensional stimuli since it can be computed using diffusion models [9] trained on natural images.ResultsWe apply that method to simulated responses from a biologically realistic model of the retina and quantify the amount of information attributed to individual pixels of natural images. These decompositions scaled well to natural images and large neural populations. We also found that encoded information is concentrated along object edges, which are features that affect the neuron's uncertainty about the stimulus the most. In contrast to Fisher information, our method accounts for both the sensitivity of the response of a neuron to small perturbation of a stimulus (its local sensitivity) and the statistics of natural stimuli (approximated by the diffusion models), revealing how neural tuning and stimulus context jointly shape encoding. DiscussionWe introduced a tractable decomposition of sensory neural information. A central hypothesis in sensory neuroscience is that neurons efficiently encode information about natural inputs given resource limits [10]. Testing this theory has been difficult due to the computational demands of measuring information in neural populations responding to complex natural stimuli [11,12]. Our method overcomes this challenge, enabling empirical tests of efficient coding via comparisons of the information encoded in real neural responses and predictions of optimal models. This framework will support new insights into how both biological and artificial neural networks represent information. Funded by Agence Nationale de la Recherche (ANR) RetNet4EC - https://doi.org/10.1002/j.1538-7305.1948.tb01338.x
- https://doi.org/10.1162/089976600300015259
- https://doi.org/10.1088/0954-898X_10_4_303
- https://doi.org/10.1007/978-1-4612-0919-5_16
- https://doi.org/10.1038/s41583-021-00502-3
- https://doi.org/10.1088/0954-898X_14_2_301
- https://doi.org/10.1371/journal.pbio.0040092
- https://doi.org/10.1007/s00422-017-0729-7
- https://doi.org/10.48550/arXiv.2006.11239
- https://doi.org/10.7551/mitpress/9780262518420.003.0013
- https://doi.org/10.1515/znc-1981-9-1040
- https://doi.org/10.3109/0954898X.2011.638888
