TITLE: Neural-Symbolic Systems for Cognitive Reasoning
SPEAKER: Artur d'Avila Garcez (School of Informatics, City University London)

ABSTRACT:

Human cognition involves the integration of reasoning and learning abilities. However, reasoning and learning are normally studied separately in computer science and artificial intelligence. In our research, we seek to integrate these abilities into neural-symbolic systems and offer a unified approach to robust learning with expressive reasoning within the neural-computation paradigm.  

In neural-symbolic systems, a neural network offers a parallel machine for computation, inductive learning and efficient reasoning, while high-level logical representations of the machine offer rigour, modularity and explanation to the network implementation. 

In this talk, I review the work on neural-symbolic systems, starting from logic programming, which has already provided contributions to problems in bioinformatics and engineering. I then look at how modal logic and other forms of non-classical reasoning can be implemented in the network model. 

Network ensembles, each representing the knowledge of an agent at a time point, can be combined at different levels of abstraction  to form modular, deep networks. These implement various reasoning tasks, including temporal, epistemic, intuitionistic, abductive, relational and uncertainty reasoning. 

Recently, the neural-symbolic model has been applied to the integrated verification and adaptation of software models, to online reasoning and learning in driving simulators and to the recognition and explanation of actions in video. The results indicate that the model is capable of controlling the accumulation of errors which is a common problem in such uncertain domains. 

This talk is based on the book "Neural-symbolic Cognitive Reasoning", Springer 2009.