σ-promoter classification
CS6024: Algorithmic Approaches to Computational Biology @ IIT Madras
The success of CNN models can be attributed to the presence of a large amounts of labelled data which allows us to train very complex models on these data and predict the labels (be it for classification, segmentation, detection or any other problem). The availability of the labelled data is critical to train huge models that work extremely well on these tasks. The same task (for example classification of digits or of objects) can require different data for different entities or individuals based on the domain of the images in their test use case. But, gathering such annotated data for every new task would be challenging, expensive and time consuming. So, it is worthwhile to look at techniques where we can adapt a classifier trained to classify images in one domain (source domain) to classify images of a different domain (target domain). This would help in transferring the learned knowledge from source to target domain, and at the same time, reduces the annotation work for target domain. The authors of the paper that we implement try to combat this problem by using an intermediate shared domain which facilitates the conversion between the source and the target domain.