Hybrid weightless neural systems

Hybrid systems are important when considering the varied nature of application domains in Artificial Intelligence (AI). Many complex domains have many different component problems, each of which requiring different types of processing. For example, Artificial Neural Networks (ANNs) have been successfully applied to solve low-level cognitive tasks such as perception, motor control and associative information retrieval. However, considerably less experience has been gathered so far in modeling high-level cognitive tasks using ANNs and on how they represent and reason about the acquired knowledge. In contrast, the ability to provide users with higher level explanations of the reasoning process is an important feature of AI systems.

Explanation facilities are required both for user acceptance of the solutions generated by the system and for the purpose of understanding whether the reasoning procedure is sound. Thus, in order for ANNs to achieve a wider degree of user acceptance and to enhance their overall utility as learning and generalization tools, it is highly desirable, if not essential, that an explanation capability becomes an integral part of the functionality of a trained ANN.

Based on this, we work on the development of a hybrid system in which symbolic rules could be inserted/extracted into/from an ANN model. More specifically, the neural model we use is the Weightless Neural Networks (WNNs). We use automata theory in the hybrid system to deal with the insertion and extraction of rules. The basic idea of such a hybrid system, called the hybrid weightless neural networks, is to insert a set of symbolic rules into a WNN. Next, the WNN could be refined by using standard WNN learning algorithms and a set of training examples. The refined network can then function as a highly accurate classifier. A final step for this system is the extraction of refined, comprehensible rules from the trained WNN.

Main publications:

• Marcilio C. P. de Souto, Jose C. M. Oliveira, and Teresa B. Ludermir. A tool to implement probabilistic automata in RAM-based neural networks. In IJCNN, pages 1054–1060, 2011. doi: 10.1109/IJCNN.2011.6033339
• Teresa B. Ludermir, Marcílio C. P. de Souto, and Wilson Rosa de Oliveira. On a hybrid weightless neural system. International Journal of Bio-Inspired Computation, 1(1/2):93–104, 2009. doi: 10.1504/IJBIC.2009.022778
• Marcílio C. P. de Souto, Teresa B. Ludermir, and Wilson Rosa de Oliveira. Equivalence between RAM-based neural networks and probabilistic automata. IEEE Transactions on Neural Networks and Learning Systems, 16(4):996–999, 2005. doi: 10.1109/TNN.2005.849838
• Marcílio C. P. de Souto, Paulo Adeodato, and Teresa B. Ludermir. Sequential RAM-based neural networks : Learnability, generalisation, knowledge extraction, and grammatical inference. International Journal of Neural Systems, 9(3):203–210, 1999. doi: 10.1142/S0129065799000198