When ultrafast laser pulses are tightly focused inside a transparent medium, various nonlinear optical phenomena readily occur at the focal point. These phenomena at the focal point can be utilized for nonlinear-optical microscopy, nanosurgery inside biological samples, device fabrication inside transparent materials, and welding of transparent materials. We review these new techniques from a unified viewpoint.

The basics of three-dimensional (3-D) and spectral imaging techniques that are based on the detection of coherence functions and other related techniques are reviewed. The principle of the 3-D source retrieval is based on understanding the propagation law of optical random field through the free space. The 3-D and spectral information are retrieved from the cross-spectral density function of optical random field or numerical calculation of the inverse propagation of the cross-spectral density. We will first introduce the coherence-based spectral tomography techniques with low-coherence light sources. These techniques limit their scheme of coherence detection only along the optical axis and some of them achieve simultaneously the high resolution and high speed of detection taking advantage of an imaging lens. We then provided explanations of the principle of 3-D source retrieval that is based on the propagation law of optical random field through the free space along with the introduction of the numerical holography and computed tomography techniques. We will lastly show 3-D spectral imaging schemes with the concurrent laser-scanning cross-sectioning techniques: one is the confocal laser scanning microscopy and the other is the two-photon laser-scanning fluorescence microscopy.

Describes various techniques to implement optical dynamic interconnections using the photorefractive effect. Two categories of optical dynamic interconnections are discussed. One is based on free-space optics and the other on guided optics. These include a real-time volume hologram, a mutually pumped phase conjugator, a spatial light modulator, a waveguide, and a spatial soliton.

In training neural networks, if a training set is not randomly extracted from the whole input domain, it is difficult to construct well generalized neural networks. In this paper, to solve the problem in image processing of corneal endothelium, we propose a method to get a well generalized shift-invariant neural network by designing optimal weights between the input and first hidden layers. The neural network has weight patterns between each layer, and its signal propagation is calculated by discrete convolution between the weight patterns and data in lower layers. The optimal weights between the input and first hidden layer are designed by observing the training set distribution in the vector space. The designed weights transform the training set to cover the whole input domain. In the learning process, the designed weights are fixed and the other weights are updated. In the experiment, well generalized neural networks are obtained.

Abstract In this paper, we propose a hyper-spherical surface interactive interconnections (HSII) learning method for category learning. This method generates the number of hidden units based on the same algorithm as a restricted coulomb energy (RCE) network and updates weights and biases by the error bbackpropagation algorithm (EBP). The HSII method adapts desired features of RCE and EBP methods. The HSII (1) optimizes the network size, (2) learns samples quickly, and (3) has well generalized performance on untrained inputs. We demonstrate application of the HSII to recognition of characters on X-ray films. We also compared performances of the HSII with the RCE using receiver operating characteristic analysis.