Shunji Satoh

Computational theory and applications of a filling-in process at the blind spot

A mathematical model for filling-in at the blind spot is proposed. The general scheme of the standard regularization theory was used to derive the model deductively. First, we present the problems encountered with a diffusion equation, which is frequently used for various types of perceptual completion. To solve these problems, we investigated the computational meaning of a neural property discovered by Matsumoto and Komatsu [Matsumoto, M., & Komatsu, H. (2005). Neural responses in the macaque V1 to bar stimuli with various lengths presented on the blind spot. Journal of Neurophysiology, 93, 2374-2387]. Based on our observations, we introduce two types of curvature information of image properties into the a priori knowledge of missing images in the blind spot. Moreover, two different information pathways for filling-in, which were suggested by results of physiological experiments (slow conductive paths of horizontal connections in V1, and fast feedforward/feedback paths via V2), were considered theoretically as the neural embodiment of an adiabatic approximation between V1 and V2 interaction. Numerical simulations show that the output of the proposed model for filling-in is consistent with neurophysiological experimental results. The model can be used as a powerful tool for digital image inpainting.

Engineering-approach accelerates computational understanding of V1–V2 neural properties

We present two computational models (i) long-range horizontal connections and the nonlinear effect in V1 and (ii) the filling-in process at the blind spot. Both models are obtained deductively from standard regularization theory to show that physiological evidence of V1 and V2 neural properties is essential for efficient image processing. We stress that the engineering approach should be imported to understand visual systems computationally, even though this approach usually ignores physiological evidence and the target is neither neurons nor the brain.

A Next Generation Modeling Environment PLATO: Platform for Collaborative Brain System Modeling

To understand the details of brain function, a large scale system model that reflects anatomical and neurophysiological characteristics needs to be implemented. Though numerous computational models of different brain areas have been proposed, these integration for the development of a large scale model have not yet been accomplished because these models were described by different programming languages, and mostly because they used different data formats. This paper introduces a platform for a collaborative brain system modeling (PLATO) where one can construct computational models using several programming languages and connect them at the I/O level with a common data format. As an example, a whole visual system model including eye movement, eye optics, retinal network and visual cortex is being developed. Preliminary results demonstrate that the integrated model successfully simulates the signal processing flow at the different stages of visual system.

Pattern Recognition System with Top-Down Process of Mental Rotation

A new model which can recognize rotated, distorted, scaled, shifted and noised patterns is proposed. The model is constructed based on psychological experiments in a mental rotation. The model has two types of processes: (i) one is a bottom-up process in which pattern recognition is realized by means of a rotation-invariant neocognitron and a standard neocognitron and (ii) the other is a top-down process in which a mental rotation is executed by means of a model of associative recall in visual pattern recognition. In computer simulations, it is shown that the model can recognize rotated patterns without training those patterns.

Parallel Numerical Simulation of Visual Neurons for Analysis of Optical Illusion

Detailed mechanism of optical illusion caused by visual neurons in human brain has not been well understood, and its numerical simulation is helpful to analyze visual system of humans. This paper describes implementation techniques of parallel numerical simulation to help understanding optical illusion by using a GPU-accelerated PC cluster. Our parallel acceleration techniques include following three points. Firstly, input images of the numerical simulation is efficiently calculated by dividing it images for multiple computation nodes using MPI (Message Passing Interface). Secondly, convolution, which is dominated computation for the optical flow, is accelerated by GPU. Finally, an algorithm to compute convolution specified to analyze optical illusion is proposed to speed up the simulation. Our experimental results show an interesting insight that values of optical flow for images causing optical illusion are quite different compared to that does not cause the optical illusion. We also demonstrate that our implementation of simulation works efficiently on the GPU-accelerated PC cluster.

Computational identity between digital image inpainting and filling-in process at the blind spot

Digital image inpainting (DII) is a computer algorithm that restores missing information of images such as those of old oil paintings. This problem occurs in human visual systems as well: we have blind spots (BS), but we see natural patterns in the BS region. This article presents the computational identity between the DII algorithm and the vision model for the filling-in process at the BS. Based on physiological evidence and conjecture, we define an evaluation function that evaluates the quality of filled-in (or inpainted) images. The definition of the evaluation function helps the original DII algorithm to improve the convergence speed. Numerical experiments demonstrate that the convergence speed using the energy function is three times faster than the original DII algorithm. Results show that the resultant filled-in patterns by the visual model are comparable with those of the DII algorithm.

Multi-GPU Acceleration of Optical Flow Computation in Visual Functional Simulation

Numerical simulation for visual processing of the human brain is one of time-consuming applications. This paper shows acceleration techniques for a simulation program of the visual processing. We parallelize convolution calculations, which are core operations, which the simulation program requests, on a GPU-accelerated PC cluster. Our implementation includes three improvement points. Firstly, we consider efficient data mapping onto global and shared memories1 of the GPU. Secondly, multiple convolutions for the same input data are computed by each nodes GPU, referred to as package execution. Finally, an input 2-dimensional image is divided into regions and convolutions for these regions are executed in parallel utilizing MPI (Message Passing Interface). Our experimental results show a linear speedup up to 12 nodes in the PC cluster for the convolution program. We also show the effects of the package execution and reduced communication on NVIDIA tesla C1060 and C2070, respectively.

Neural implementation of coarse-to-fine processing in V1 simple neurons

Coarse-to-fine processing has been observed in various areas of the visual cortex. For example, some receptive fields (RFs) of neurons in the primary visual cortex (V1) shrink spatially as time progresses. Such V1 neurons become more sensitive to higher spatial period stimulus in 20ms. Furthermore, orientation selectivity in V1 also increases, that is, orientational coarse-to-fine processes. As described herein, we investigate the neural substances related to coarse-to-fine processing. We find that such coarse-to-fine processing corresponds to deblurring operations to achieve V1 neural output with spatially and orientationally higher resolutions. We show computationally that the short-range horizontal connections (SHCs) realize the deblurring operation. We simulate a V1 network model with SHCs based on the Sasaki model [1] (Sasaki and Satoh, 2009). The shrinking V1 receptive-field and increased orientational selectivity are caused by neural deblurring operations through SHCs. The model properties are qualitatively consistent with physiological data.

Super resolution: Another computational role of short-range horizontal connection in the primary visual cortex

Recent physiological data related to the primary visual cortex (V1) have shown various contextual effects in the non-classical receptive field (nCRF). Contextual modulation, size tuning and altered sensitivity of orientation are typical examples of such contextual effects in the nCRF. These phenomena in the nCRF have been thought to be caused by short-range horizontal connection (SHC). However, SHC does not necessarily contribute only to these phenomena. These phenomena might be merely secondary phenomena by the fundamental role of SHC. In this paper, we specifically address the overcomplete properties in V1. Then the fundamental role of SHC is examined from image-processing points of view. Super resolution is proposed as a strong candidate for the fundamental role of SHC. Super resolution is an engineering method that obtains a high-resolution image from a low-resolution image(s). The distribution of SHC is deductively derived by adopting a reverse diffusion technique, which is one of various available super-resolution techniques. The spatial distribution of our proposed SHC is isotropic on the orientation map. This characteristic is consistent with physiological data. In addition to that, contextual modulation, size tuning and altered sensitivity of orientation in numerical experiments using our proposed SHC can be reproduced qualitatively. These results indicate that these phenomena are secondary phenomena by super-resolution processing.

Formulation of Border-Ownership Assignment in Area V2 as an Optimization Problem

Border-ownership (BO) assignment, or the assignment of borders to an occluding object, is a primary step in visual perception. Physiological experiments have revealed the existence of neurons in area V2 that respond selectively to objects placed on a specific side of their response field. Although existing models can reproduce this phenomenon, they are not based on a clear computational theory. For this study, we formulated BO assignment as a well-defined optimization problem. We hypothesize that information related to BO assignment can be expressed as a conservative vector field. This conservative vector field is proposed as the gradient of a scalar field that carries information related to the depth order of the overlapping object. Conservative vector fields have zero curl (rotation). Using this theorem, we construct and solve an optimization problem. Numerical simulations demonstrate that a model based on our derived algorithm solves BO assignment for problems of perceived order and occlusion. Deduced neural networks provide insight into possible characteristics of lateral connections in area V2.