Plenary Lecture

Neural Networks: A Bridge Towards Self-Observation

Associate Professor Jean-Jacques Mariage
Computing Science Department
Laboratoire d' Informatique Avancee de Saint-Denis (LIASD)
2, Rue de la Liberte
93526 Saint-Denis Cedex, France
E-mail: jam@ai.univ-paris8.fr

Abstract: Regardless of their increasing number and diversity, the capacities of Neural Network (NN) models still remain far behind the ones biological systems can exhibit when faced to changing environments or other complex processes. As an attempt to better understand why, we propose to investigate the variations of a few NN algorithms in the theoretical framework of Darwinian evolution in order to favor the emergence of more global models through gradual adaptive developmental steps.
First, we explore the possibility to achieve a more general conception of learning and training methodology, detached from specialized NN models. Then, with the ultimate goal to bypass human bias constraints in data acquisition, we apply NNs to the automatic categorization of natural language data without prior knowledge.
To this end, we argue that enhanced plasticity and reorganization capabilities are necessary for NNs in order to be able to detect and structurally integrate variations in the data space. A first step is to model and simulate the dynamic character of the “biological” learning structures and processes as well as their evolution over time. We propose a dual architecture, where two –possibly identical– NNs collaborate, one learning to control the efficiency of the other. This way, a reflexive loop of self-supervision can be achieved where one NN learns to tune the configuration parameters (wiring, growth, learning rate, etc.) of the other through automated trial and error sessions. A further step is the use of data driven programming combined with error measures in the self-supervision loop to create a self-observing retroactive loop in order to analytically develop an active, event guided, learning. The previously mentioned dual architecture can then be used to learn to extract and apply characteristic learning features of other NN models. NNs would thereby, in response to environmental changes, put into practice their acquired adaptive developmental capabilities to generate the appropriate variations, both at the architectural and procedural level.
We will also distinguish different scales of structural variations in the natural – and mostly biological – world in order to illustrate emergent steps in the evolution of developmental strategies, similar to those nature has selected.

Brief Biography of the Speaker:
Jean-Jacques Mariage was born in Saisseval, France, in 1953. He is Associate Professor of Computer Science at the University of Paris 8 since 2001, where he joined the Artificial Intelligence Laboratory. He teaches at the Franco Georgian Institute since 2006. His teaching activity involves Artificial Intelligence, programming languages and computer network engineering. His current research addresses the integration of a self-observing retroactive loop in unsupervised Neural Network (NN) models, applied to the automatic categorization of natural language data without previous knowledge. To this end, his interest focuses on the modeling and simulation of the developmental dynamics of adaptive encoding structures as found in biological systems. His areas of interest include learning, memory, evolution, NN algorithms, genetic algorithms, evolutionary programming, artificial life as well as the development, replication and adaptation of biological encoding structures.