EVE Optimiser Interactive Figures

Hello World! Thank you for visiting the website. You are in the right place! This page is the secret sauce of our paper. The main aim of this page is to enable the community to explore the proposed EVE optimisation method performance in more grandular detail as interactive figures. I hope you find the website helpful. Please feel free to reach out if you have any questions/feedback.

Usage

• Double click on the legend items to isolate individual learning rate trajectories. To reset simply double click once again.
• Head to the top right corner of each figure to explore more utilities. For example, you will be able to zoom in an area of interest in the figure by using the Zoom tool.

Figure 1

Figure Description

Illustration of the training loss under a range of primary learning rate variations. As shown, the marginal histogram of the training loss illustrates the superiority of EVE over Adam on CIFAR-100 data.

Figure 2

Figure Description

Illustration of the training loss under a range of primary and second learning rate variations for EVE. As shown, the various values for primary and second learning rate lead to the same performance according to the marginal training loss distribution with the majority of the settings hitting below 0.2999 as the training loss value.

Figure 3

Figure Description

Illustration of the validation losses for EVE and Adam using CIFAR-100 dataset. As shown, EVE outperforms Adam by achieving consistent validation loss values during the training which will yield improved generalisation capacity.

Figure 4

Figure Description

Illustration of the training loss under a range of primary learning rate variations. As shown, the marginal histogram of the training loss illustrates the superiority of EVE over Adam on Flower Classification Data.

Figure 5

Figure Description

Illustration of the training F1-score trajectories under a range of primary learning rate variations. As shown, the marginal histogram of the training F1-score illustrates the superiority of EVE over Adam on Flower Classification data.

Figure 6

Figure Description

Illustration of the validation loss trajectories under a range of primary learning rate variations. As shown, the marginal histogram of the validation loss illustrates the superiority of EVE over Adam on Flower Classification data.

Figure 7

Figure Description

Illustration of the validation F1-score trajectories under a range of primary learning rate variations. As shown, the marginal histogram of the validation F1-score illustrates the superiority of EVE over Adam on Flower Classification data.

Figure 8

Figure Description

Illustration of the training and validation loss trajectories under a range of primary and second learning rate variations for EVE on Flower Classification data. As shown, the various values for primary and second learning rate lead to the same performance according to the marginal training and validation loss distributions with the majority of the settings hitting below 0.05 and 0.3 for the training and validation loss values, respectively.

Figure 9

Figure Description

Illustration of the training and validation F1-score trajectories under a range of primary and second learning rate variations for EVE on Flower Classification data. As shown, the various values for primary and second learning rate lead to the same performance according to the marginal training and validation F1-score distributions with the majority of the settings.