Pseudo-recurrent Connectionist Networks: An Approach to the 'Sensitivity-Stability' Dilemma

摘要: In order to solve the 'sensitivity-stability' problem-and its immediate correlate, problem of sequential learning-it is crucial develop connectionist architectures that are simultaneously sensitive to, but not excessively disrupted by, new input. French (1992) suggested alleviate a particularly severe form this disruption, catastrophic forgetting, it was necessary for networks separate dynamically their internal representations during learning. McClelland et al. (1995) went even further. They nature's way implementing obligatory separation evolution two areas brain, hippocampus and neocortex. keeping with idea radical separation, 'pseudo-recurrent' memory model presented here partitions network into functionally distinct, continually interacting areas. One area serves as final-storage representations; other an early-processing where firs...