José Luis Crespo

Training Higher Order Gaussian Synapses

In this article we present an algorithm that permits training networks that include gaussian type higher order synapses. This algorithm is an extension of the classical backpropagation algorithm. Higher order synapses permit carrying out tasks using simpler networks than traditionally employed. The key to this simplicity is in the structure of the synapses: a gaussian with three trainable parameters. The fact that it is a function and consequently presents a variable output depending on its inputs and that it possesses more than one trainable parameter that allows it to implement non linear processing functions on its inputs, endows the networks with a large capacity for learning and generalization. We present two examples where these capacities are shown. The first one is a target tracking module for a the visual system of a real robot and the second one is an image classification system working on real images.

A new image prediction model based on spatio-temporal techniques

This paper addresses an image prediction problem focused on images with no identifiable objects. In it, we present several approaches to predict the next image of a given sequence, when the image lacks the well-defined objects, such as meteorological maps or satellite imagery. In these images no clear borders are present, and any object candidate moves, changes, appears and disappears in any image. Nevertheless, this evolution, though unrestricted, is gradual and, hence, prediction looks feasible. One of the approaches presented here, based on a spatio-temporal autoregressive (STAR) model, offers good results for these kinds of images.The main contribution of this paper is to adapt spatio-temporal models to an image prediction problem.As a byproduct of this research, we have achieved a new image compression method, suitable for images without defined shapes.

An adaptive detection/attention mechanism for real time robot operation

During the lifetime of a mobile robot, the number and complexity of the stimuli it receives may be quite high. Therefore, the construction of a detection system considering the whole sensorial space is usually not a viable proposition when aiming for real time operation. It becomes necessary to build some kind of sensorial hierarchy map to put some order into how detectors are applied. This is what is usually called an attentional system, and it provides a framework for applying detectors in a more efficient manner. In this paper, an architecture for developing attentional functions for robots that must operate in real time in dynamic environments is presented. This architecture is based on the concept of attentor and it allows for the real time adaptation to the environment and tasks to be performed in a natural manner. One of the main requirements imposed on the design of the architecture was the capability of handling different sensorial modalities and attentional streams in a transparent manner while, at the same time, being able to progressively create more complex attentional structures. The architecture is particularized for its implementation in a real robot.

Adaptation, Performance and Vapnik-Chervonenkis Dimension of Straight Line Programs

We discuss here empirical comparation between model selection methods based on Linear Genetic Programming. Two statistical methods are compared: model selection based on Empirical Risk Minimization (ERM) and model selection based on Structural Risk Minimization (SRM). For this purpose we have identified the main components which determine the capacity of some linear structures as classifiers showing an upper bound for the Vapnik-Chervonenkis (VC) dimension of classes of programs representing linear code defined by arithmetic computations and sign tests. This upper bound is used to define a fitness based on VC regularization that performs significantly better than the fitness based on empirical risk.

Preprocessing Phase in the PIETSI Project (Prediction of Time Evolution Images Using Intelligent Systems)

We outline the PIETSI project, the core of which is an image prediction strategy, and discuss common preliminary image processing tasks that are relevant when facing problems such as: useless background, information overlapping (solvable if dissimilar coding is being used) and memory usage, which can be described as “marginal information efficiency”.

Moving objects forecast in image sequences using autoregressive algorithms

The objective of this paper is to present an overall approach to forecasting the future position of the moving objects of an image sequence after processing the images previous to it. The proposed method makes use of classical techniques such as optical flow to extract objects’ trajectories and velocities, and autoregressive algorithms to build the predictive model. Our method can be used in a variety of applications, where videos with stationary cameras are used, moving objects are not deformed and change their position with time. One of these applications is traffic control, which is used in this paper as a case study with different meteorological conditions to compare with.

Meteorological image descriptors

The objective of this paper is to get a visual characterization of time evolution images, in particular, synoptic maps taken from Meteorology. Preliminary tasks required before image processing are reviewed. Two different types of numerical descriptors are extracted for characterizing the images, the called low level numerical descriptors, and the high level corresponding ones. The latter will be subsequently used for prediction tasks, meanwhile the former will be used for classification tasks. Three different relevant information sources in the images are identified as their low level descriptors. These are defined by the local density and orientation of the isobar lines, and the number of centres of high (H) and low (L) pressure. Regarding the high level descriptors, two main features are taken into account. The different procedures carried out to extract the previous descriptors for our images of interest are discussed.

Gaussian Synapse Networks for Hyperspectral Image Segmentation

In this work we have made use of a new type of network with non linear synapses, Gaussian Synapse Networks, for the segmentation of hyperspectral images. These structures were trained using the GSBP algorithm and present two main advantages with respect to other, more traditional, approaches. On one hand, through the intrinsic filtering ability of the synapses, they permit concentrating on what is relevant in the spectra and automatically discard what is not. On the other, the networks are structurally adapted to the problem as superfluous synapses and/or nodes are implicitly eliminated by the training procedure.

Tests of Different Regularization Terms in Small Networks

Several regularization terms, some of them widely applied to neural networks, such as weight decay and weight elimination, and some others new, are tested when applied to networks with a small number of connections handling continuous variables. These networks are found when using additive algorithms that work by adding processors. First the different methods and their rationale is presented. Then, results are shown, first for curve fitting problems. Since the network constructive algorithm is being used for system modeling, results are also shown for a toy problem that includes recurrency buildup, in order to test the influence of the regularization terms in this process. The results show that this terms can be of help in order to detect unnecessary connections. No clear winner has been found among the presented terms in these tests.

Sensibility Analysis of an Object Movement Forecast Approximation in Real Image Sequences

The objective of this paper is to analyse the influence of the different parameters used for an overall approach to forecasting a future position of the mobile objects of an image sequence after processing the previous images to it. Our approximation uses classical techniques such as optical flow to extract objects trajectories and velocities and autoregressive algorithms to build the predictive model. Applications to outdoor scenarios are possible, for videos where stationary cameras are used and moving objects follow an affine displacement field. In this work, traffic sequences with different meteorological conditions are studied.