Portal de Periódicos da CAPES

Semantic-based regularization and Piaget’s cognitive stages

2009; Elsevier BV; Volume: 22; Issue: 7 Linguagem: Inglês

10.1016/j.neunet.2009.06.048

1879-2782

Marco Gori,

Fuzzy Logic and Control Systems

ARE COGNITIVE STEPS A SECRET TO TO BREAK THE COMPLEXITY OF LEARNING? In the last decades, the focus on the research in machine learning has hovered pendulum-like from biologically inspired to artificial models with different degree of cognitive plausibility. As the time goes by and research evolves, we will likely see men and machines facing an increasingly number of similar learning tasks. Now, regardless of the extent to which they share biological principles, one might be interested in studying human and artificial cognitive processes under the same umbrella. In this paper we claim that the principles of cognitive development stages, that have been the subject of an in-depth analysis in children by Jean Piaget [1, 2] are likely to inspire important advances in machine learning. He pointed out that we can identify four major stages or periods of development in child learning, where each stage is self-contained and builds upon the preceding stage. In addition, children seem to proceed through these stages in a universal, fixed order. They start developing sensorimotor and preoperational skills, in which the perceptual interactions with the environment dominate the learning process, and evolve by exhibiting concrete and formal operational skills, in which they start to think logically and develop abstract thoughts. When observing human and nowadays artificial minds on the same play, one early realizes that machines do not take into account most of the rich human communication protocols. In most of the studies of machine learning, the agent is expected to learn from labelled and/or un-labelled examples finalized to a specific task. There are, however, a number of other crucial interactions of the agent that are rarely taken into account. Human learning experiences witness the importance of asking questions and of learning under a of teaching plan. While the first interaction has been considered in a number of machine learning models, apart from a remarkable exception [3], to the best of our knowledge, teaching plans have not been significantly involved in learning algorithms. What is often neglected in machine learning is that most intriguing human learning skills are due, to a large extent, to the acquisition of relevant semantic attributes and to their relations. This makes learning a process which goes well beyond pure induction; the evidence provided by the induction of a semantic attribute is typically propagated to other attributes by formal rules, thus giving rise to a sort of reinforcement cyclic process. ∗When I was in high school, my physics teacher whose name was Mr. Bader called me down one day after physics class and said, “you look bored; I want to tell you something interesting. Then he told me something which I found absolutely fascinating, and have, since then, always found fascinating. Every time the subject comes up, I work on it. Richard Feynman, in physics lectures, on the principle of least action.