next up previous contents index
Next: Learning algorithms for DTRNN Up: Sequence processing with neural Previous: Other architectures without hidden   Contents   Index


Application of DTRNN to sequence processing

DTRNN have been applied to a wide variety of sequence-processing tasks; here is a survey of some of them:

Channel equalization:
In digital communications, when a series of symbols is transmitted, the effect of the channel may yield a signal whose decoding may be impossible without resorting to a compensation or reversal of these effects at the receiver side. This sequence transduction task (which converts the garbled sequence received into something as similar as possible to the transmitted signal) is usually known as equalization. A number of researchers have studied DTRNN for channel equalization purposes (Ortiz-Fuentes and Forcada, 1997; Bradley and Mars, 1995; Parisi et al., 1997; Cid-Sueiro and Figueiras-Vidal, 1993; Kechriotis et al., 1994; Cid-Sueiro et al., 1994).
Speech recognition:
Speech recognition may be formulated either as a sequence transduction task (for example, continuous speech recognition systems aim at obtaining a sequence of phonemes from a sequence of acoustic vectors derived from a digitized speech sample) or as a sequence recognition task (for example, as in isolated-word recognition, which assigns a word in a vocabulary to a sequence of acoustic vectors). Discrete-time recurrent neural networks have been extensively used in speech recognition tasks (Watrous et al., 1990; Chiu and Shanblatt, 1995; Robinson, 1994; Kuhn et al., 1990; Bridle, 1990; Chen et al., 1995; Robinson and Fallside, 1991).
Speech coding:
Speech coding aims at obtaining a compressed representation of a speech signal so that it may be sent at the lowest possible bit rate. A family of speech coders are based in the concept of predictive coding: if the speech signal at time $t$ may be predicted using the values of the signal at earlier times, then the transmitter may simply send the prediction error instead of the actual value of the signal and the receiver may use a similar predictor to reconstruct the signal; in particular, a DTRNN may be used as a predictor. The transmission of the prediction error may be arranged in such a way that the number of bits necessary is much smaller than the one needed to send the actual signal with the same reception quality (Sluijter et al., 1995). Haykin and Li (1995), Baltersee and Chambers (1997), and Wu et al. (1994) have used DTRNN predictors for speech coding.
System identification and control:
DTRNN may be trained to be models of time-dependent processes such as a stirred-tank continuous chemical reactor: this is usually referred to as system identification. Control goes a step further: a DTRNN may be trained to drive a real system (a ``plant'') so that the properties of its output follows a desired temporal pattern. Many researchers have used of DTRNN in system identification (Nerrand et al., 1994; Werbos, 1990; Adali et al., 1997; Cheng et al., 1995; Dreider et al., 1995), and control (Chovan et al., 1996; Narendra and Parthasarathy, 1990; Chovan et al., 1994; Puskorius and Feldkamp, 1994; Zbikowski and Dzielinski, 1995; Wang and Wu, 1995; Li et al., 1995; Wang and Wu, 1996).
Time series prediction:
The prediction of the next item in a sequence may be interesting in many other applications besides speech coding. For example, short-term electrical load forecasting is important to control electrical power generation and distribution. Time series prediction is a classical sequence prediction application of DTRNN. See, for example, Draye et al. (1995); Connor and Martin (1994); Aussem et al. (1995); Dreider et al. (1995).
Natural language processing:
The processing of sentences written in any natural (human) language may itself be seen as a sequence processing task, and has been also approached with DTRNN. Examples include discovering grammatical and semantic classes of words when predicting the next word in a sentence (Elman, 1991), learning to assign thematic roles to parts of Chinese sentences (Chen et al., 1997), or training a DTRNN to judge on the grammaticality of natural language sentences (Lawrence et al., 1996).


next up previous contents index
Next: Learning algorithms for DTRNN Up: Sequence processing with neural Previous: Other architectures without hidden   Contents   Index
Debian User 2002-01-21