What question did this study set out to answer?

To develop a deep learning framework capable of effectively training neural networks on datasets with noisy or incorrect labels by explicitly estimating the noise process.

March 1, 2016

Training deep neural-networks based on unreliable labels

View Full Paper

Ask AI

Bookmark

Authors

ABAlan Joseph BekkerBar-Ilan University JGJacob GoldbergerBar-Ilan University

Discussion

Member takes

Implication

Computational study demonstrates improved classification performance by modeling label noise channels, indicating neural networks can learn effectively from corrupted datasets.

Key Points

To develop a deep learning framework capable of effectively training neural networks on datasets with noisy or incorrect labels by explicitly estimating the noise process.
Added an extra noise layer representing an unknown noisy channel that converts true labels into observed labels.
Trained the model to simultaneously learn neural network weights and the underlying noise distribution, benchmarking against standard back-propagation on several standard classification tasks.
Demonstrated improved classification performance over standard training pipelines that do not account for incorrect labels.
Showed performance benefits even on manually annotated datasets that were presumed to be clean and error-free.

Cite This Study

Bekker et al. (2016) studied this question.

synapsesocial.com/papers/6a11dbe48095bbd48eb4d636 https://doi.org/10.1109/icassp.2016.7472164

View Full Paper

Ask AI

Bookmark

Implication

Computational study demonstrates improved classification performance by modeling label noise channels, indicating neural networks can learn effectively from corrupted datasets.

Training deep neural-networks based on unreliable labels

Discussion

Member takes

Implication

Key Points

Cite This Study

Training deep neural-networks based on unreliable labels

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider