Potter J R, Mellinger D K, Clark C W
Scripps Institution of Oceanography, University of California San Diego, La Jolla 92093-0238.
J Acoust Soc Am. 1994 Sep;96(3):1255-62. doi: 10.1121/1.410274.
Recent work has applied a linear spectrogram correlator filter (SCF) to detect bowhead whale (Balaena mysticetus) song notes, outperforming both a time-series-matched filter and a hidden Markov model. The method relies on an empirical weighting matrix. An artificial neural net (ANN) may be better yet, since it offers two advantages; (i) the equivalent weighting matrix is determined by training and can converge to a more optimal solution and (ii) an ANN is a nonlinear estimator and can embody more sophisticated responses. A three-layer feed-forward ANN is ideally suited to this application and has been implemented on 1475 sounds, of which 54% were used for training and 46% kept as "unseen" test data. The trained ANN error rate was 1.5%, a twofold improvement over previous methods. It is shown that ANN hidden neurons can be interrogated to reveal the operating paradigm developed during training. The function of each of these neurons can be determined in terms of spectrographic features of the training calls. Furthermore, the operating paradigm can be controlled and training time reduced by assigning specific recognition tasks to hidden neurons prior to training, rather than initiating training with randomized weights. The ANN is compared to the SCF and the role of the "hidden" neurons and equivalent weighting matrices are discussed.
最近的研究工作应用了线性谱图相关器滤波器（SCF）来检测弓头鲸（Balaena mysticetus）的歌声音符，其性能优于时间序列匹配滤波器和隐马尔可夫模型。该方法依赖于一个经验加权矩阵。人工神经网络（ANN）可能更胜一筹，因为它具有两个优点：（i）等效加权矩阵通过训练确定，并且可以收敛到更优的解决方案；（ii）人工神经网络是非线性估计器，可以体现更复杂的响应。三层前馈人工神经网络非常适合此应用，并且已经在1475个声音上实现，其中54%用于训练，46%留作“未见”测试数据。训练后的人工神经网络错误率为1.5%，比以前的方法提高了两倍。结果表明，可以对人工神经网络的隐藏神经元进行询问，以揭示训练过程中形成的操作范式。这些神经元中的每一个的功能都可以根据训练叫声的频谱特征来确定。此外，通过在训练前为隐藏神经元分配特定的识别任务，而不是用随机权重启动训练，可以控制操作范式并减少训练时间。将人工神经网络与SCF进行了比较，并讨论了“隐藏”神经元和等效加权矩阵的作用。
