Neural net architectures for temporal sequence processing

I present a general taxonomy of neural net architectures for processing time-varying patterns. This taxonomy subsumes many existing architectures in the literature, and points to several promising architectures that have yet to be examined. Any architecture that processes timevarying patterns requires two conceptually distinct components: a short-term memory that holds on to relevant past events and an associator that uses the short-term memory to classify or predict. My taxonomy is based on a characterization of short-term memory models along the dimensions of form, content, and adaptability. Experiments on predicting future values of a financial time series (US dollar–Swiss franc exchange rates) are presented using several alternative memory models. The results of these experiments serve as a baseline against which more sophisticated architectures can be compared. Neural networks have proven to be a promising alternative to traditional techniques for nonlinear temporal prediction tasks (e.g., Curtiss, Brandemuehl, & Kreider, 1992; Lapedes & Farber, 1987; Weigend, Huberman, & Rumelhart, 1992). However, temporal prediction is a particularly challenging problem because conventional neural net architectures and algorithms are not well suited for patterns that vary over time. The prototypical use of neural nets is in structural pattern recognition. In such a task, a collection of features—visual, semantic, or otherwise—is presented to a network and the network must categorize the input feature pattern as belonging to one or more classes. For example, a network might be trained to classify animal species based on a set of attributes describing living creatures such as “has tail”, “lives in water”, or “is carnivorous”; or a network could be trained to recognize visual patterns over a two-dimensional pixel array as a letter in {A,B, . . . , Z}. In such tasks, the network is presented with all relevant information simultaneously. In contrast, temporal pattern recognition involves processing of patterns that evolve over time. The appropriate response at a particular point in time depends not only on the current input, but potentially all previous inputs. This is illustrated in Figure 1, which shows the basic framework for a temporal prediction problem. I assume that time is quantized into discrete steps, a sensible assumption because many time series of interest are intrinsically discrete, and continuous series can be sampled at a fixed interval. The input at time t is denoted x(t). For univariate series, this input