Dear Hritam Basak

These articles might useful, have a look:

1. https://hal.archives-ouvertes.fr/hal-03049937/file/Ouali_Semi-Supervised_Semantic_Segmentation_With_Cross-Consistency_Training_CVPR_2020_paper.pdf

2. Preprint Semi-supervised Medical Image Segmentation through Dual-task...

3. Preprint Domain Adaptive Video Segmentation via Temporal Consistency ...

4. Article Deep Neural Network With Consistency Regularization of Multi...

5. Article Transformation-Consistent Self-Ensembling Model for Semisupe...

Kind Regards

Qamar Ul Islam

Thanks, Qamar Ul Islam for your response.

However, I'd like to know why there is the necessity of consistency regularization, and why is it better than a single encoder-decoder network? Any thoughts, rather than a paper would be more useful.

Dear Hritam Basak,

In a caricatural way, conventional deep learning tends to mimic rote learning, i.e. it generally provides very satisfactory answers to patterns for which it has precisely learnt, but it can unexpectedly provides completely wrong answers to patterns that are very close to correctly learnt patterns.

This problem is referred to as a generalization problem:

• Zhang et al., "Understanding Deep Learning (Still) Requires Rethinking Generalization", 2021 - https://cacm.acm.org/magazines/2021/3/250713-understanding-deep-learning-still-requires-rethinking-generalization/fulltext

In order to solve this generalization problem, the solutions that are mostly explored nowadays are based on the assumption that any learning problem can be modeled by means of a low-dimensional manifold. In order to have all the chances to discover this manifold in its generality with a restricted set of training data, the "Transformation Consistency Regularization" method consists in intentionally perturbing in an amplified way the input learning patterns. In this way, it is hoped to encourage a variant of learning that is less based on "rote learning".

In my opinion, this kind of method tends to approach some topological concerns in a heuristic way.

I suggest you the following paper which presents in a didactic way the topic of "Transformation Consistency Regularization":

• Mustafa et al, "Transformation Consistency Regularization-A Semi-Supervised Paradigm for Image-to-Image Translation", 2020 - https://www.cl.cam.ac.uk/~rkm38/pdfs/mustafa2020_transformation_consistency_regularization.pdf

Regards