AN EFFECTIVE NEURAL MODEL MECHANIZING HARD CAUSAL REASONING PROBLEMS WITH WTA and WTO NEURAL COMPUTATIONS

In this work, we develop a neural model to solve causal reasoning problems (said also abduction) in the open, independent and incompatibility classes. We model the reasoning process by a single and global energy function using cooperative and competitive neural computation. The update rules of the distinct connections of the network are derived from its energy function using gradient descent techniques. Simulation results reveal a good performance of the model.

Max-product algorithms for the generalized multiple-fault diagnosis problem

In this paper, we study the application of the max-product algorithm (MPA) to the generalized multiple-fault diagnosis (GMFD) problem, which consists of components (to be diagnosed) and alarms/connections that can be unreliable. The MPA and the improved sequential MPA (SMPA) that we develop in this paper are local-message-passing algorithms that operate on the bipartite diagnosis graph (BDG) associated with the GMFD problem and converge to the maximum a posteriori probability (MAP) solution if this graph is acyclic (in addition, the MPA requires the MAP solution to be unique). Our simulations suggest that both the MPA and the SMPA perform well in more general systems that may exhibit cycles in the associated BDGs (the SMPA also appears to outperform the MPA in these more general systems). In this paper, we provide analytical results for acyclic BDGs and also assess the performance of both algorithms under particular patterns of alarm observations in general graphs; this allows us to obtain analytical bounds on the probability of making erroneous diagnosis with respect to the MAP solution. We also evaluate the performance of the MPA and the SMPA algorithms via simulations, and provide comparisons with previously developed heuristics for this type of diagnosis problems. We conclude that the MPA and the SMPA perform well under reasonable computational complexity when the underlying diagnosis graph is sparse.

A softmin-based neural model for causal reasoning

This paper extends a neural model for causal reasoning to mechanize the monotonic class. Hence, the resulting model is able to solve multiple, varied causal problems in the open, independent, incompatibility and monotonic classes. First, additivity between causes is formalized as a fuzzy AND-ing process. Second, an activation mechanism called the "softmin" is developed to solve additive interactions. Third, the softmin is implemented within a neural architecture. Experimental results on real-world and artificial problems reveal a good performance of the model and should stimulate future research.

AN ARTIFICIAL NETWORK FOR REASONING IN THE CANCELLATION CLASS WITH APPLICATION TO THE DIAGNOSIS OF CELLS DIVISION

Causal reasoning is a hard task that cognitive agents perform reliably and quickly. A particular class of causal reasoning that raises several difficulties is the cancellation class. Cancellation occurs when a set of causes (hypotheses) cancel each other's explanation with respect to a given effect (observation). For example, a cloudy sky may suggest a rainy weather; whereas a shiny sky may suggest the absence of rain. In this work we extend a recent neural model to handle cancellation interactions. Simulation results are very satisfactory and should encourage research.

X-tron: an incremental connectionist model for category perception

A connectionist model for categorization (self-organization) even in the presence of multiple or mixed patterns has been presented. During self-organization, the network automatically adjusts the number of nodes in the hidden and output layers, depending on the complexity or nature of overlap between the patterns. An ambiguity measure is given based on how well the features are being interpreted by the network. From the ambiguity measure a certainty factor about the decision of the network is derived. The effect of noise on the certainty factor is investigated. A vigilance threshold is used to decide whether the network's decision is correct or not. Functionally the network consists of two parts, one of them categorizes the incoming patterns and the other monitors the performance of categorization. The characteristics of the model has also been demonstrated experimentally on both 1D binary strings and image patterns even when they are corrupted by additive, subtractive, and mixed noise.

High-specificity neurological localization using a connectionist model

Most previous connectionist models for diagnosis have been developed using error backpropagation. While these systems function reasonably well, they have been limited by their need for a large database of test cases, to situations where a single disorder is present, and by the large number of connections required between fully-connected sets of processing units. Here we describe a recently developed connectionist model that overcomes these limitations. This approach can reuse existing causal knowledge bases, works well in situations where multiple disorders can occur simultaneously, and does not require fully-connected sets of processing units. We demonstrate that the accuracy of this model is comparable to that of more conventional AI programs using the same knowledge base in determining precisely the site of brain damage in a group of 50 stroke patients. These results support the conclusion that connectionist models can effectively use pre-existing causal knowledge bases from AI systems, and that they can function accurately when handling actual clinical problems.

Connectionist expert systems as medical decision aid

In this paper neural networks are used as associative memories to build an expert system for aiding medical diagnosis. As in expert systems using symbolic manipulation, the knowledge is introduced by a knowledge engineer using a collection of known cases. The system has an object-oriented approach to knowledge organization and the resulting network topology. Fuzzy sets are used to interpret connection values and/or excitation state of the units. The main result is that the proposed neural network allows not only finding a solution in some cases, but also suggests obtaining more clinical data if the data available is insufficient to reach a conclusion. This approach is illustrated by examples.

Neural computation in medicine

During the last decade there has been a great revival of interest in neural modelling. Powerful new computational methods have resulted from work in this area and are being applied to an increasing range of medical problems. This paper briefly explains the nature of a neural model and then reviews work in neural computation involving problems in medical informatics (e.g. expert systems) and modelling of psychiatric and neurological phenomena. The state of the art is assessed, and speculation about future developments is given.

Phase Transitions in Connectionist Models Having Rapidly Varying Connection Strengths

A phase transition in a connectionist model refers to a qualitative change in the model's behavior as parameters determining the spread of activation (gain, decay rate, etc.) pass through certain critical values. As connectionist methods have been increasingly adopted to model various problems in neuroscience, artificial intelligence, and cognitive science, there has been an increased need to understand and predict these phase transitions to assure meaningful model behavior. This paper extends previous results on phase transitions to encompass a class of connectionist models having rapidly varying connection strengths (“fast weights”). Phase transitions are predicted theoretically and then verified through a series of computer simulations. These results broaden the range of connectionist models for which phase transitions are identified and lay the foundation for future studies comparing models with rapidly varying and slowly varying connection strengths.