GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 3.1 GNGTS 2023 As we can appreciate from the figure, we note that the application of the CL model on real data results in a segmentation map in which many more pixels are activated than in the single training case. Figure 3: Results of predictions conducted on real data by both models: CL corresponds to the predictions of the model trained with Curriculum Learning strategy, while 'single' stands for the model trained on random complexity instances. Both predictions are compared with the results of a deterministic algorithm for LWD automated interpretation, based on Sobel filter and Dynamic Time Warping workflow, referred in the figure to as ‘Petropick’. The results shown here are the average predictions of the networks on 1 meter sliding window of real data. We can say that CL results demonstrate how fine-tuning the network with increasing complexity of the instances can generate predictions of higher-level features, with respect to a “single” training with no fine-tuning and random inistances with random complexity during fitting.