An Application of the Saturated Attractor Analysis to Three Typical Models
Jianfeng Feng,B. Tirozzi
DOI: https://doi.org/10.1007/3-540-59497-3_196
1995-01-01
Abstract:The saturated attractor analysis, an approach proposed first in (FP) for a comprehensive study of the dynamics of the Linsker model and then successfully applied to the dynamic link model(FT1), is further developed here. By a unified approach to the Hopfield model, the Linsker model and the dynamic link model, three typical models in the field of the neural networks, we show a way to choose the parameters of these dynamics in order to obtain any chosen saturated attractor which is general enough in most applications. We generalize our previous results for the Linsker model and the dynamic link model with the clipping function to the case of the sigmoid like function. Our results allow us for the first time to understand the underlying mechanism among these models and thus to furnish a useful guidance in the further possible applications. spin glas~ theory to the Hopfield model allows us to understand clearly the phase transition from the retrieval to non retrieval state. Another major impulse was provided by the successful explanation of some biological phenomena, at least in a primitive level, for example, the Linsker model mimics the ontogenesis development of the primary visual system(Lin). Of course, the most important impulse comes from the learning techniques successfully applied to some practical problems which were traditionally thought of as some of the hardest problems in the AI. One of the recent examples of such an application is the face recognition using the dynamic link model, a model proposed by vonder Malsbnrg first in 1981(KMM). However, at this moment, the theoretical treatment of these models is obviously far away from being satisfactory, mainly due to the lack of theoretical tools to deal with the nonlinearity exploited in most of the models reported today. In the present paper, in terms of our previous work on the Linsker model and the dynamic link model we develop a unified theoretical framework for tackling the Hopfield model, the Linsker model and the dynamic link model. Our approach allows us to reformulate many problems studied in the Hop field model. A concrete criterion to check whether a stored pattern is an attractor of the network is given. The capacity, a quantity which plays a central role in the spin glass approach to the Hopfield model, is naturally introduced here. One advantage of the present approach is that we do not impose the restriction of the symmetry of the connection matrix. Our results also reveal the role of different parameters in the Hopfield model. We consider the Linsker model with the sigmoid like function in the updating dynamics of its synaptic connections(a definition of the slgmoid like function is in section 2). All conclusions in (FP)(FPR) are reobtalned, where the clipping function, a special case of the sigmoid like function and so a special case of the present paper, is used for the development of the synaptic connections. The present paper tells that the appearance of the structured receptive fields is independent of the choice of the clipping function, which is thought of as one of the drawback of the Linsker model. Furthermore, we also take into account on the reason for the appearance of the oriented receptive field in the further layers of the Linsker network. For the dynamic link model, a principle to choose all five parameter employed in the model is furnished, which confirms our previous claim that all results contained in (FT1) for the clipping function are true for a more general class of function, i.e. for tile sigmoid like function. Although here we confine ourselves to the models on which we worked before (ATYD) (FP) (FPP,) (FQ) (FT1), the essential part of our approach is to analyze the dynamics with the sigmoid like function, and