Add documentation for the noisy label detection feature (#2034)

* Add documentation for noisy label detection feature

Signed-off-by: Kim, Vinnam <[email protected]>

* Update CHANGELOG.md

Signed-off-by: Kim, Vinnam <[email protected]>

* Add documentation

Signed-off-by: Kim, Vinnam <[email protected]>

* Fix typo

Signed-off-by: Kim, Vinnam <[email protected]>

* Fix small typo

---------

Signed-off-by: Kim, Vinnam <[email protected]>
Co-authored-by: Songki Choi <[email protected]>

Author: vinnamkim

CHANGELOG.md

@@ -11,6 +11,7 @@ All notable changes to this project will be documented in this file.
 - Support export of onnx models (<https://github.com/openvinotoolkit/training_extensions/pull/1976>)
 - Add option to save images after inference in OTX CLI demo together with demo in exportable code (<https://github.com/openvinotoolkit/training_extensions/pull/2005>)
 - Support storage cache in Apache Arrow using Datumaro for cls, det, seg tasks (<https://github.com/openvinotoolkit/training_extensions/pull/2009>)
+- Add noisy label detection for multi-class classification task (<https://github.com/openvinotoolkit/training_extensions/pull/1985>, <https://github.com/openvinotoolkit/training_extensions/pull/2034>)
 
 ### Enhancements
 

docs/source/guide/explanation/additional_features/index.rst

@@ -10,3 +10,4 @@ Additional Features
    hpo
    auto_configuration
    xai
+   noisy_label_detection

docs/source/guide/explanation/additional_features/noisy_label_detection.rst

@@ -0,0 +1,27 @@
+Noisy label detection
+=====================
+
+OpenVINO™ Training Extensions provide a feature for detecting noisy labels during model training.
+With this feature, you can identify noisy labeled samples in your training dataset.
+Our algorithm accumulates the training loss dynamics during the model training
+and exports it to `Datumaro <https://github.com/openvinotoolkit/datumaro>`_.
+The training loss dynamics are then post-processed by exponential moving average (EMA),
+a strong criterion for detecting noisy label samples [1]_.
+Finally, Datumaro ranks the top-k samples, which can be considered as noisy labeled candidates.
+We provide an end-to-end example written in a Jupyter notebook, which you can find at the link in the note below.
+
+In OpenVINO™ Training Extensions CLI, you can enable this feature
+by adding the argument ``--algo_backend.enable_noisy_label_detection true`` as follows.
+
+.. code-block::
+
+    $ otx train ... params --algo_backend.enable_noisy_label_detection true
+
+.. note::
+    Currently, it only supports multi-class classification task and single GPU training.
+
+.. note:: **Important!**
+    The post-processing step to analyze the training loss dynamics requires `Datumaro <https://github.com/openvinotoolkit/datumaro>`_.
+    Please see `this end-to-end Jupyter notebook example <https://github.com/openvinotoolkit/datumaro/blob/develop/notebooks/10_noisy_label_detection.ipynb>`_.
+
+.. [1] Zhou, Tianyi, Shengjie Wang, and Jeff Bilmes. "Robust curriculum learning: from clean label detection to noisy label self-correction." International Conference on Learning Representations. 2021.