Learning Curves Prediction for a Transformers-Based Model

One of the main challenges when training or fine-tuning a machine learning model concerns the number of observations necessary to achieve satisfactory performance. While, in general, more training observations result in a better-performing model, collecting more data can be time-consuming, expensive, or even impossible. For this reason, investigating the relationship between the dataset's size and the performance of a machine learning model is fundamental to deciding, with a certain likelihood, the minimum number of observations that are necessary to ensure a satisfactory-performing model is obtained as a result of the training process. The learning curve represents the relationship between the dataset’s size and the performance of the model and is especially useful when choosing a model for a specific task or planning the annotation work of a dataset. Thus, the purpose of this paper is to find the functions that best fit the learning curves of a Transformers-based model (LayoutLM) when fine-tuned to extract information from invoices. Two new datasets of invoices are made available for such a task. Combined with a third dataset already available online, 22 sub-datasets are defined, and their learning curves are plotted based on cross-validation results. The functions are fit using a non-linear least squares technique. The results show that both a bi-asymptotic and a Morgan-Mercer-Flodin function fit the learning curves extremely well. Also, an empirical relation is presented to predict the learning curve from a single parameter that may be easily obtained in the early stage of the annotation process. Doi: 10.28991/ESJ-2023-07-05-03 Full Text: PDF