Parallel Studies on Artificial Neural Networks and Vertebrate Retinal Neurosystems

(1)鲤鱼视网膜短波锥体向H_1水平细胞(H_1HC)的突触传递是以APB为激动剂的电导减低型突触。突触为外侧视网膜的信号传递提供了一种新的可塑性机制。这一机制可以解释为什么光适应条件下H1HC感受野在蓝/绿光刺激下比红色刺激下更窄的古老问题。(2)中心-周围对手感受野(CSRF)通过边缘检测学习在3-1层反向传播神经网络(NN)中出现。CSRF结构以这样一种方式改变，以提高噪声破坏的输入模式中的边缘的可检测性。这种依赖SN的可塑性在真实的视觉系统中也存在，在果蝇复眼中存在视网膜双极细胞和某些中间神经元。(3)部分应用了(1)的可塑性机制，提出了一种用于视觉图像处理的自适应空间滤波器。构建了一个自联想记忆网络，并基于压缩映射原理对其进行了分析。发现存在大的吸引域，并且在这样的域中的任何噪声模式都可以被识别为存储的模式之一。该网络已成功地应用于字符和彩色图像的识别问题。(5)设计了一种新的算法来动态消除神经网络中不必要的连接。在性能误差被最小化的同时，可以通过对每个突触的输入突触权重参数应用相互抑制规则来消除神经连接的冗余。

(1) The synaptic transmission from short-wavelength cones to H1 horizontal cells (H1HCs) in the carp retina is found to be of a conductance-decrease type with APB as an agonist. The synapse provides a new plasticity scheme for the lateral signal transmission across the outer retina. This mechanism can explain the old question of why the H1HC receptive field under light-adapted condition is narrower with blue/green lights than with red stimuli.(2) Center-surround opponent receptive fields (CSRF) emerge in a 3-1ayerbackpropagation neural network (NN), as a result of edge-detection learning. The CSRF structure changes in such a way as to increase detectability of edges in the noise-corrupted input patterns. The same kind of SN-dependent plasticity exists in real visual systems, with retinal bipolar cells as well as certain interneurons in the fly compound eyes.(3) An adaptive spatial filter for visual image processing has been developed, in part by applying the plasticity mechanisms of (1). The VLSI fabricated operates to solve a regularization problem automatically.(4) An autoassociative memory network is constructed and it is analyzed on the basis of the contraction mapping principle. Large domains of attraction are found to exist, and any noisy pattern in such a domain can be recognized as one of the stored patterns. The network has been applied successfully to character and color-picture recognition problems.(5) A new algorithm is devised to eliminate dynamically unnecessary connections in NN. While the performance error is being minimized, redundancy of neural connections can be removed by applying a mutual suppressionrule over input synaptic weight parameters at each synapse.