Connectionist Temporal Classification

E736823

Connectionist Temporal Classification is a neural network training algorithm designed for sequence labeling tasks where input and output lengths differ and alignments are unknown, widely used in speech and handwriting recognition.

Try in SPARQL Jump to: Statements Referenced by

Statements (48)

Predicate Object
instanceOf loss function
neural network training algorithm
sequence labeling method
abbreviation CTC NERFINISHED
advantage enables end-to-end training
no need for pre-aligned training data
assumes conditional independence between output labels given network outputs
basedOn recurrent neural networks
category probabilistic sequence model
comparedWith HMM-based sequence labeling
coreIdea sums over all valid alignments between input and output sequences
designedFor sequence labeling tasks
field deep learning
handwriting recognition
machine learning
speech recognition
handles unsegmented input data
variable-length input sequences
variable-length output sequences
hasComponent alignment paths
collapse function from paths to label sequences
inspired end-to-end ASR systems such as Deep Speech
introducedBy Alex Graves NERFINISHED
Faustino Gomez NERFINISHED
Jürgen Schmidhuber NERFINISHED
Santiago Fernández NERFINISHED
lossFamily negative log-likelihood loss
objective maximize log-likelihood of correct label sequence
oftenUsedWith RNN acoustic models
bidirectional LSTM networks
optimizationMethod gradient-based optimization
outputType label sequence probabilities
publicationYear 2006
publishedIn Proceedings of the 23rd International Conference on Machine Learning NERFINISHED
relatedTo attention mechanisms
sequence-to-sequence models
requires a special blank label
solves alignment-free sequence labeling
supports training without frame-level alignments
trainingSignal sequence-level supervision
usedIn end-to-end speech recognition
offline handwriting recognition
online handwriting recognition
optical character recognition
scene text recognition
sign language recognition
uses dynamic programming
forward-backward algorithm

Referenced by (1)

Full triples — surface form annotated when it differs from this entity's canonical label.

Alex Graves notableWork Connectionist Temporal Classification