Easy incremental learning methods to consider for commercial fine-tuning applications

Udhaya Ravishankar

Keywords: Incremental Learning, Neural Networks, Classification, Regression

TL;DR: The paper proposes a few simplified methods to consider, for implementing incremental learning in commercial fine-tuning infrastructures.

Abstract: Fine-tuning deep learning models for commercial use cases is growing exponentially as more and more companies are adopting AI to enhance their core products and services, as well as automate their diurnal processes and activities. However, not many countries like the U.S. and those in Europe follow quality data collection methods for AI vision or NLP related automation applications. Thus, on many of these kinds of data, existing state-of-the-art pre-trained deep learning models fail to perform accurately, and when fine-tuning is done on these models, issues like catastrophic forgetting or being less specific in predictions as expected occur. Hence, in this paper, simplified incremental learning methods are introduced to be considered in existing fine-tuning infrastructures of pre-trained models (such as those available in huggingface.com) to help mitigate the aforementioned issues for commercial applications. The methods introduced are: 1) Fisher Shut-off, 2) Fractional Data Retention and 3) Border Control. Results show that when applying these methods on vanilla pre-trained models, the models are in fact able to add more to their knowledge without hurting much on what they had learned previously.