Neuromorphic Retina: An FPGA-based Emulator

TL;DR

The paper tackles the challenge of creating a biologically plausible retinal model suitable for prosthetics by implementing a neuromorphic digital retina on an FPGA. It introduces a Convis-inspired architecture with a center-surround outer plexiform layer (OPL) using non-separable spatio-temporal filters, luminance adaptation via a high-pass photoreceptor stage, a bipolar-level contrast gain control, and an inner plexiform layer (IPL) with leaky integrate-and-fire (LIF) ganglion cells that generate spikes. The OPL computes $I_OPL = λ_OPL (C - ω_OPL S)$ with $C = G_C * T_{w,τ} * E_{τC} * L$ and $S = G_S * E_{τS} * C$. The bipolar gain control is described by $C dV_{Bip}/dt = I_{OPL} - g_A V_{Bip}$ and $g_A = G_A * E_A * Q(V_{Bip})$, with $Q(V_{Bip}) = g^0_A + λ_A V_{Bip}^2$, enabling adaptive luminance/contrast processing and tunable phasic/tonic responses; the implementation supports 128×128 input at 200fps on an Artix-7 and demonstrates real-time operation for retinal prosthesis applications.

Abstract

Implementing accurate models of the retina is a challenging task, particularly in the context of creating visual prosthetics and devices. Notwithstanding the presence of diverse artificial renditions of the retina, the imperative task persists to pursue a more realistic model. In this work, we are emulating a neuromorphic retina model on an FPGA. The key feature of this model is its powerful adaptation to luminance and contrast, which allows it to accurately emulate the sensitivity of the biological retina to changes in light levels. Phasic and tonic cells are realizable in the retina in the simplest way possible. Our FPGA implementation of the proposed biologically inspired digital retina, incorporating a receptive field with a center-surround structure, is reconfigurable and can support 128*128 pixel images at a frame rate of 200fps. It consumes 1720 slices, approximately 3.7k Look-Up Tables (LUTs), and Flip-Flops (FFs) on the FPGA. This implementation provides a high-performance, low-power, and small-area solution and could be a significant step forward in the development of biologically plausible retinal prostheses with enhanced information processing capabilities

Figures (11)

• Figure 1: \ref{['retina_model']} Simplified schema showing the retina implementation in the digital realm, starting from photoreceptor cell receiving external video stimuli from the outside environment to the spike response generated by the ganglion cells; \ref{['software_op']} Images depicting the effect of stimulus at different retina layer model where, (1) shows the $10^{th}$ frame of an input video stimulus, (2) shows the impact of the OPL layer output, (3) illustrates the result of the bipolar layer, and (4) shows the resultant spiking activity.
• Figure 2: Kernels used in the neuromorphic digital retina for spatial and temporal filtering in the retinal layers; \ref{['spatialfilter']} Depicts spatial low-pass filter kernel (Gaussian) for $G_\text{C}$, $G_\text{S}$, and $G_\text{A}$; \ref{['lowpass']} Shows temporal low-pass filter kernel for $E_{\tau \text{C}}$, $E_{\tau \text{S}}$, and $E_{\tau \text{A}}$; \ref{['highpass']} Shows temporal high-pass filter kernel for $T_{\text{w},\tau}$ and $T_{G}$. Symbol $\sigma$ indicates the spatial constant, while $\tau$ represents the time constant.
• Figure 3: \ref{['qfuncmarch03']} Shows the activation function, denoted as $Q(V_\text{Bip})$ for the synaptic conductances $g_{A}$ associated with bipolar level contrast adaptation. This implies that the activation of $g_{A}$ is solely determined by the absolute value of $V_\text{Bip}$, the bipolar potential; \ref{['gi_res']} Shows the function that rectifies signals from bipolar to ganglion cells, denoted as $N(V_\text{Bip})$. It is a reflection of static nonlinearities that have been empirically discovered in the retinachichilnisky2001simplebaccus2002fastkim2001temporal.
• Figure 4: The digital architecture of \ref{['ooplarch']} the outer plexiform layer (OPL) accepts video frames of size 128$\times$128 as input for its photoreceptor cells and generates corresponding output frames called $I_\text{OPL}$, also with a size of 128$\times$128; \ref{['bipoarchi']} Depicts that the pink-shaded region represents a bipolar layer incorporating contrast gain control. It accepts $I_\text{OPL}$ as the input frame with a size of 128$\times$128, from the OPL layer and generates an output frame of the same size, 128$\times$128, known as $V_\text{Bip}$ which is gain controlled. Excitatory current $I_\text{Gang}$ is generated within the blue-shaded area and serves as the input to a LIF neuron responsible for spike generation.
• Figure 5: \ref{['convblock']} Shows the block diagram of convolution operation for spatial filtering in retinal layers. \ref{['conv33']} Shows the architecture of $3\times3$ register bank for spatial filtering in the center signal components of the OPL layer. \ref{['conv55']} Shows the architecture of $5\times5$ register bank for spatial filtering in the surround signal components of the OPL layer, and bipolar level contrast gain control stage.