Variational studies and replica symmetry breaking in the generalization problem of the binary perceptron
Carregando...
Data
2000-11-01
Autores
Botelho, Evaldo
Mattos, Cristiano R. [UNESP]
Caticha, Nestor
Título da Revista
ISSN da Revista
Título de Volume
Editor
Resumo
We analyze the average performance of a general class of learning algorithms for the nondeterministic polynomial time complete problem of rule extraction by a binary perceptron. The examples are generated by a rule implemented by a teacher network of similar architecture. A variational approach is used in trying to identify the potential energy that leads to the largest generalization in the thermodynamic limit. We restrict our search to algorithms that always satisfy the binary constraints. A replica symmetric ansatz leads to a learning algorithm which presents a phase transition in violation of an information theoretical bound. Stability analysis shows that this is due to a failure of the replica symmetric ansatz and the first step of replica symmetry breaking (RSB) is studied. The variational method does not determine a unique potential but it allows construction of a class with a unique minimum within each first order valley. Members of this class improve on the performance of Gibbs algorithm but fail to reach the Bayesian limit in the low generalization phase. They even fail to reach the performance of the best binary, an optimal clipping of the barycenter of version space. We find a trade-off between a good low performance and early onset of perfect generalization. Although the RSB may be locally stable we discuss the possibility that it fails to be the correct saddle point globally. ©2000 The American Physical Society.
Descrição
Palavras-chave
Computer simulation, Entropy, Failure analysis, Gibbs free energy, Integration, Learning algorithms, Monte Carlo methods, Neural networks, Phase transitions, Polynomials, Potential energy, Thermodynamic stability, Binary perceptrons, Replica symmetry breaking (RSB), Statistical mechanics
Como citar
Physical Review E - Statistical Physics, Plasmas, Fluids, and Related Interdisciplinary Topics, v. 62, n. 5 B, p. 6999-7007, 2000.