摘要：A common belief in the community is that deep learning requires large datasets to be effective. We show that with careful parameter selection, deep feature extraction can be applied even to small datasets.We also explore exactly how much data is necessary to guarantee learning by convergence analysis and calculating the shattering coefficient for the algorithms used. Another problem is that state-of-the-art results are rarely reproducible because they use proprietary datasets, pretrained networks and/or weight initializations from other larger networks. We present a two-fold novelty for this situation where a carefully designed CNN architecture, together with a knowledge-driven classifier achieves nearly state-of-the-art phoneme recognition results with absolutely no pretraining or external weight initialization. We also beat the best replication study of the state of the art with a 28% FER. More importantly, we are able to achieve transparent, reproducible frame-level accuracy and, additionally, perform a convergence analysis to show the generalization capacity of the model providing statistical evidence that our results are not obtained by chance. Furthermore, we show how algorithms with strong learning guarantees can not only benefit from raw data extraction but contribute with more robust results.
关键词：feature extraction; acoustic modeling; phoneme recognition; statistical learning theory
