visionfans的专栏

Sequence training is short for sequence discriminative training.ReferenceVoigtlaender, Paul, et al. “Sequence-discriminative training of recurrent neural networks.” Acoustics, Speech and Signal Proces

The Impact of Imbalanced Training Data for Convolutional Neural Networks Following this, oversampling is used on the imbalanced training sets to increase the performances to that of the balanced set.

Viterbi training has another name segmental K-means.Within each state, a further alignment of observations to mixture components is made. HTK provides two methods for this: 1. use clustering to alloc

M. Mohri, F. C. N. Pereira, and M. Riley. Weighted FiniteState Transducers in Speech Recognition. Computer Speech and Language, 16(1):69–88, 2002.Allauzen, C., Mohri, M., Riley, M., Roark, B., A Gener

Data matrix XX shape (D,N)(D, N). Zero mean centering or not depending on circumstance. Compute eigen vectors with SVD: U,S,V=linalg.svd(X)U, S, V = \texttt{linalg.svd}(X)Projection: X′=U⊺∗XX' = U

For natural images, it make little sense to have separate mean and variance values for each pixel. They are already have equivalent variances.The desiderata are that our learning algorithms sees a tra

ASR evaluationTranscription accuracyWER(Word Error Rate)SER(Sentence Error Rate)Semantic accuracyConcept Accuracy(based on domain concepts)

A weighted finite state transducer (WFST) T(x,y)T(x, y) over a semiring is defined as:T(x,y)=⊕π∈P(I,x,y,F)λ(p[π])⊗w[π]⊗ρ(n[π])T(x, y) = \underset{\pi\in P(I,x,y,F)}{\oplus}\lambda(p[\pi])\otimes w[\p

ARPA format language modelhttp://www.statmt.org/book/slides/07-language-models.pdfToolkits: - SRILM - IRSTLM

Key concepts:tf-idfinverted indexReferencesHow a Search Engine Works Information Retrieval: What is inverted index?Implementing Search Engines

cross entropy or log lossIn binary classification, the output is a single probability value yy. L(x,z)=−(zln(y)+(1−z)ln(1−y))\mathcal{L}(x, z)=-(z\ln(y)+(1-z)\ln(1-y))For multi-class problem, the ou

2 road lines: - EM algorithm(Baum-Welch) - Gradient DescentThe main advantage of the Baum-Welch algorithm (and hence the ML training) is due to its simplicity and the fact that requires no parameter

Nomial