[Blog #3] Six Weeks at LBNL

Six weeks out, my learning curve became steeper. My conceptual understanding of Transfer Learning really improved, and I understood its future potential in the Machine Learnin Industry. My current task was almost coming to an end - recording information about literature pertaining to the use of Transfer Learning in energy efficiency in buildings. In total, our aim is to go through 75 odd papers as part of our literature review, so I have myself gone through arounnd 30 papers. This has developed my understanding of general ideas, like research writing, conventions used in research papers, how to read a paper quickly to understand its salient factors and findings. The main difficulty was understanding the type of Transfer Learning (TL) used in the context of saving energy, as there are several sub-categories.

First category: whether the TL used is Inductive or Transductive. Essentially, TL aims to use the information from a particular source (say, a group of buildings) and use the knowledge learned in the Machine Learning framework for a target (like a target building without too much data on thermal fluctuations, occupancy, load, etc. that would be required to train the ML models to save energy). Sometimes, the TL used learns to predict a certain item (like load) for the source domain, but it's trained in such a way that it develops generalized knowledge and can be used to predict some other item (like occupancy). This is Inductive TL (similar to Inductive Logic). However, if the source and target tasks are the same, it's Transductive TL.

Second category: whether the TL is Homogeneous or Heterogeneous. If two buildings have the same input features (temperature, load, cooling power, etc.), then the task is Homogeneous. If the TL is such that the model trained on the source building uses 5 features, but it has to use only 2 features on the target, then the task is Heterogeneous.

There's a third category: Instance-based, Feature-based, Parameter-based, and Relational-based. Each have their own nuances, and it took time for me to truly get a good understanding of each category. All in all, it's been amazing, I've got to learn so much by reading so many papers. The next step is to start writing my designated section in the Review paper that will be published, so I'm excited.