Neuromorphic models for biological photoreceptors

Abstract

Biological visual processing is extremely flexible and provides pixelby- pixel adaptation. Millennia of evolution and natural selection have provided inspiration for robust, efficient and elegant solutions in artificial visual system designs. Physiological studies have shown that non-linear adaptation of biological visual processing is evident even at the first stage of the visual system pathway. Theory and modelling have shown that adaptation in the early visual processing is required to compress the high bandwidth visual environment into a sensible form prior to transmission via the limited bandwidth neuron channels. However, many current bio-inspired visual systems have neglected the importance of having a reliable early stage of visual processing. Having a robust and reliable early stage design not only provides a better mimic of the biology, but also allows better design and understanding of higher order neurons in the visual system pathway. (Chapter 3: A Non-linear Adaptive Artificial Photoreceptor Circuit - Design and Implementation) The primary aim of this work was to design and implement an elaborated artificial photoreceptor circuit which faithfully mimics the actual biological photoreceptors, using standard analogue discrete electronic components. I have incorporated several key features of the biological photoreceptors in the implementation, such as non-linear adaptation to background luminance, adaptive frequency response and logarithmic encoding of luminance. Initial parameters for the key features of the model were based on existing literature and fine tuning of the circuit was done after analysis of actual recordings from biological photoreceptors. (Chapter 2: Dimmable Voltage-Controlled High Current LED Driver System for Vision Science Experiments) The visual stimulus was a critical component in performing the vision experiments, and has historically been a limiting factor in performing experiments which ask critical questions about responses to complicated scenes, such as natural environments. The ability to reproduce the large dynamic range of the real-world luminance was important to correctly test the performance of the model. I evaluated the performance of several existing light emitting diode (LED) drivers and commercial products and found that none of them provided adequate dynamic range and freedom from noise. I therefore designed and implemented a stable multi-channel, high-current LED driver that allowed creation of light stimuli with inexpensive analogue discrete electronic components, and was used for the experiments described in this thesis. This LED driver, which was properly calibrated to the real-world luminance, was used in conjunction with a standard commercial data acquisition card. (An Elaborated Electronic Prototype of a Biological Photoreceptor - Steady-state Analysis (Chapter 4) & Dynamic Analysis (Chapter 5)) I performed electrophysiological experiments measuring the responses of the intact hoverfly photoreceptor cells (Rl-6) using both characterised and dynamic (naturalistic) stimuli. The analysed data were used to fine tune the circuit parameters in order to realise a faithful mimic of the actual biological photoreceptors. Similar experiments were performed on the artificial photoreceptor circuit to thoroughly evaluate the robustness and performance of the circuit against actual biological photoreceptors. Correlation and coherence analyses were used to measure the performance of the circuit with respect to its biological counterpart in both time and frequency domains respectively. Chapter 6: Early Visual Processing Maximises Information for Higher Order Neurons) The artificial photoreceptor circuit was then further evaluated against a complex natural movie scene in which the full dynamic range of the original scenario was maintained. Again, I performed experiments on both the circuit and actual biological photoreceptors. Correlation and coherence analyses of the circuit against the biological photoreceptors showed that the circuit was robust and reliable even under complex naturalistic conditions. I managed to design and implement an add-on electronic circuit to the elaborated photoreceptor circuit that crudely mimicked the temporal high-pass nature of the second order Large Monopolar Cell (LMC) in order to observe how the non-linear features in the early stage of visual processing assists higher order neurons in efficiently coding visual information. Based on this research, I found that the first stage of visual processing consists of numerous non-linearities, which have been proven to provide optimal coding of visual information. The variable frequency response curve of the hoverfly, Eristalis tenax was mapped out against large range of background luminance. Previous studies have suggested that such variability in frequency response was to improve signal transmission quality in the insect visual pathway, even though I have not made any quantitative measurements of the improvements. I also found that high dynamic range images (32-bit floating point numbers) are better representations of the real-world luminance for naturalistic visual experiments compared to the conventional 8-bit images. I have successfully implemented a circuit that faithfully mimicked the biological photoreceptors and it has been evaluated against characterised and dynamic stimuli. I found that my circuit design was far better than using just a normal linear phototransducer as the front-end of a vision system as it is more capable of compressing visual information in a way which maximises the information content before transmission to higher order neurons.