The retina is a very complex neural structure, which contains many different types of neurons interconnected with great precision, enabling sophisticated conditioning and coding of the visual information before it is passed via the optic nerve to higher visual centers. The encoding of visual information is one of the basic questions in visual and computational neuroscience and is also of seminal importance in the field of visual prostheses. In this framework, it is essential to have artificial retina systems able to function in a way as similar as possible to the biological retinas. This paper proposes an automatic evolutionary multi-objective strategy based on the NSGA-II algorithm for tuning retina models. Four metrics were adopted for guiding the algorithm in the search of those parameters that best approximate a synthetic retinal model output with real electrophysiological recordings. Results show that this procedure performs well when different trade-offs has to be considered during the design of customized neuro prostheses.

A significant challenge in neuroscience is understanding how visual information is encoded in the retina. Such knowledge is extremely important for the purpose of designing bioinspired sensors and artificial retinal systems that will, in so far as may be possible, be capable of mimicking vertebrate retinal behaviour. In this study, we report the tuning of a reliable computational bioinspired retinal model with various algorithms to improve the mimicry of the model. Its main contribution is two-fold. First, given the multi-objective nature of the problem, an automatic multi-objective optimisation strategy is proposed through the use of four biological-based metrics, which are used to adjust the retinal model for accurate prediction of retinal ganglion cell responses. Second, a subset of population-based search heuristics—genetic algorithms (SPEA2, NSGA-II and NSGA-III), particle swarm optimisation (PSO) and differential evolution (DE)—are explored to identify the best algorithm for fine-tuning the retinal model, by comparing performance across a hypervolume metric. Nonparametric statistical tests are used to perform a rigorous comparison between all the metaheuristics. The best results were achieved with the PSO algorithm on the basis of the largest hypervolume that was achieved, well-distributed elements and high numbers on the Pareto front.

Abstract. The retina is responsible for transducing visual information into spikes trains which are then sent via the optical nerve to the visual cortex. This is the first step in the visual pathway responsible for the sense of vision. Our research group is working on the design of a cortical visual neuroprosthesis aimed to restore some functional vision to profoundly visual-impaired people. The goal of developing such a bioinspired retinal encoder is not simply to record a high-resolution image, but to process its visual information and transmit it in a meaningful way to the appropriate area on the visual cortex. Retinal models to be implemented have to match as much as possible the output produced by an actual biological retina. The models involve a big search space defined by a set of parameters that have to be appropriately adjusted. This in itself has several problems which need to be addressed. We propose in this paper an automatic evolutionary multi-objective strategy for selecting those parameters which best approximate the outputs by the synthetic retina model and the biological records. A case study is presented where results of a retina model tuned with our method are compared to biological recordings.

No abstract