Gilles Vaucher

A Spatio-temporal Perceptron for On-Line Handwritten Character Recognition

The objective of this work is the application of the spatio-temporal multilayer perceptron (ST-MLP) developed in our laboratory to the recognition of on-line handwritten characters. The ST-MLP integrates a spatio-temporal data coding defined in the complex domain. Starting from the stroke of a character produced by a digitizing tablet, we conduct the recognition process in two steps. This procedure which is classic in this domain, consist of a preprocessing step and a recognition one. The first step (segmentation step), identifies some elementary (basic) lines, called primitives, from the stroke of the character. Then we utilise the ST-MLP to recognize the traced character from the primitives provided.

A complex-valued spiking machine

The technique presented in this article comes from the formalization of some passive electric properties of the dendritic trees in biological neurons using a complex-valued spatio-temporal coding. The introduction of this coding in the complex-valued neural networks makes it possible to give an algebraic formalization to the spiking neural networks. Our finality in term of applications is the processing of spatio-temporal patterns in the field of human-computer interactions. This technique was thus evaluated by simulation on handwritten character recognition and, audio and visual speech recognition problems. To improve the performances of it we present in this paper a design methodology which helps building multinetwork spiking machines. The implementation of a pen oriented interface made for an industrialist illustrates the method.

An algebraic interpretation of PSP composition.

The introduction of time in artificial neurons is a delicate problem on which many groups are working. Our approach combines some properties of biological models and the algebraic properties of McCulloch and Pitts artificial neuron (AN) (McCulloch and Pitts, 1943) to produce a new model which links both characteristics. In this extended artificial neuron, postsynaptic potentials (PSPs) are considered as numerical elements, having two degrees of freedom, on which the neuron computes operations. Modelled in this manner, a group of neurons can be seen as a computer with an asynchronous architecture. To formalize the functioning of this computer, we propose an algebra of impulses. This approach might also be interesting in the modelling of the passive electrical properties in some biological neurons.

Image sequence analysis using a spatio-temporal coding for automatic lipreading

In this paper we present the details of a processing technique utilised to extract parameters from the images of a talking persons mouth region. These parameters are converted into impulse sequences for a given series of images. Spatio-temporal (ST) coding of the impulses provide an input to a spatio-temporal neural network architecture. The system is tested on a French digit recognition task and the results obtained are given. The overall procedure is sufficiently simple to allow us to make plans for its realisation as a real-time system.

A Set of Neural Tools for Human-Computer Interactions: Application to the Handwritten Character Recognition, and Visual Speech Recognition Problems

This paper presents a new technique of data coding and an associated set of homogenous processing tools for the development of Human Computer Interactions (HCI). The proposed technique facilitates the fusion of different sensorial modalities and simplifies the implementations. The coding takes into account the spatio-temporal nature of the signals to be processed in the framework of a sparse representation of data. Neural networks adapted to such a representation of data are proposed to perform the recognition tasks. Their development is illustrated by two examples: one of on-line handwritten character recognition; and the other of visual speech recognition.

Kohonen Maps Applied to Fast Image Vector Quantization

Vector Quantization (VQ) is a powerful technique for image compression but its coding complexity may be an important drawback. Self-Organizing Maps (SOM) are well suited for topologically ordered codebook design. We propose to use that topology for reducing image coding time. Using inter-block correlations, the nearest neighbor search is restricted to the neighborhood of the precedingly used code vector instead of the entire codebook. We obtained a reduction of up to 84% in the coding time compared to full search.

Neuro-Biological Bases for Spatio-Temporal Data Coding in Artificial Neural Networks

Taking as a starting point Wilfrid Ralls dendritic tree model, well known by neuro-biologists, we propose a spatio-temporal data coding to introduce time in an artificial neuron (AN). This paper describes the biological origin of the coding and its links with the AN. With this type of coding, the algebraic properties of ANs are maintained and applied to sequence processing.

Study of a Self-Learning Artifical Neuron Model

Giving a large autonomy to each cell and using a statistical average of the entries as a self-learning mechanism are not new ideas [1,2]. But linking statistical and linear dependencies through algebric properties is an approach which leads to biological interpretations.