COMPLETE GRADIENT OPTIMIZATION OF A RECURRENT NETWORK APPLIED TO B,D,G DISCRIMINATION

被引:10
作者
WATROUS, RL
LADENDORF, B
KUHN, G
机构
[1] UNIV PENN,MOORE SCH ELECT ENGN,PHILADELPHIA,PA 19104
[2] INST DEF ANAL CRD,PRINCETON,NJ 08540
关键词
D O I
10.1121/1.398805
中图分类号
O42 [声学];
学科分类号
070206 ; 082403 ;
摘要
Two techniques for complete gradient optimization of a given connectionist network with recurrent links are described. The first technique computes network outputs on the forward pass, and contributions to the complete (or truncated) gradient on a backward pass. The second technique computes both network outputs and contributions to the complete gradient on the forward pass. The two techniques are demonstrated by being used to train the given network for talker-dependent discrimination of /b/, /d/, and /g/ in the context of a following /i/, /a/, or /u/. Consonant discrimination was as high as 98% on test syllables. © 1990, Acoustical Society of America. All rights reserved.
引用
收藏
页码:1301 / 1309
页数:9
相关论文
共 15 条
  • [1] EXPERIMENTS ON NEURAL NET RECOGNITION OF SPOKEN AND WRITTEN TEXT
    BURR, DJ
    [J]. IEEE TRANSACTIONS ON ACOUSTICS SPEECH AND SIGNAL PROCESSING, 1988, 36 (07): : 1162 - 1168
  • [2] HINTON G, 1989, COMMUNICATION
  • [3] JORDAN MI, 1986, ICS8604 I COGN SCI T
  • [4] KUHN GM, 1987, SCIMP487 I DEF AN CO
  • [5] Lippmann R. P., 1988, Computer Architecture News, V16, P7, DOI [10.1109/MASSP.1987.1165576, 10.1145/44571.44572]
  • [6] LIPPMANN RP, 1909, NEURAL COMPUT, V1, P1
  • [7] Olano C. A., 1983, Proceedings of ICASSP 83. IEEE International Conference on Acoustics, Speech and Signal Processing, P773
  • [8] Learning State Space Trajectories in Recurrent Neural Networks
    Pearlmutter, Barak A.
    [J]. NEURAL COMPUTATION, 1989, 1 (02) : 263 - 269
  • [9] ROHWER R, 1987, 1ST IEEE INT C NEUR, V2, P701
  • [10] RUMELHART DE, 1986, PARALLEL DISTRIBUTED, pCH8