This is the last week of my ten-week research journey!

As discussed last week, for the linear and JPEG data, the built-in resize method of the transforms class from torchvision was used during the experiment. However, applying similar transformations to log images disrupts the inherent relationships and yields unpredictable outcomes. Therefore, geometric resizing for log data should be conducted in linear space before transforming to log space, as resizing frequently involves pixel value interpolation – an operation best suited for linear data. After making this modification, I obtained some desired results.

Utilizing log data during both training and testing yielded the highest training accuracy at 98.36%, albeit the test accuracy is approximately 2% lower than that achieved by the linear network, which exhibited an impressive 93.8% accuracy. Notably, both networks trained with linear and log data demonstrated higher training and testing accuracy than the network trained with sRGB data. Another intriguing observation is that employing log data during both training and testing led to a higher degree of overfitting, with the testing accuracy being approximately 7% lower than the training accuracy. In comparison, the network trained with sRGB data exhibited a difference of 4%, while the linear data-trained network displayed a difference of 0.4%.

These outcomes align with the hypothesis that using linear and log data transformed from sRGB data provides advantages to neural network performance, particularly in terms of testing and training accuracy, even when the original data is already in a compressed form. However, the anticipated result that log data, which retains the most physics features, would yield the highest performance was not achieved.

In future work, overfitting could potentially be further diminished by experimenting with different optimizers (e.g., AdamW) and tuning hyperparameters. Previous research suggests that a smaller learning rate can yield more consistent log network training, and adjusting default hyperparameters (e.g., epsilon) for ADAM can aid convergence for log networks.17 While not attained in this research, I believe that resolving overfitting would likely enable log data to achieve the highest training and testing accuracy.

Many thanks to my research advisor Dr. Bruce Maxwell, my DREAM mentor Dr. Jacob Furst, and researcher Heather Fryling from Dr. Maxwell’s lab. Dr. Maxwell is always there for me for questions and advice related and beyond the research topics that I was working on. Dr. Furst always checks in with me and helped me with technical problems along the way. Heather’s code and learning tips for computer vision are really helpful for me to finish the task on time without struggling too much. I feel supported and strong in such a welcoming team. I cannot achieve huge success and satisfaction without their support and guidance.

Last but not least, thanks to DREAM, Kathleen Kelly, and Justyna Glode for providing this great opportunity for students who are career changers but would like to gain more in-depth computer science knowledge and to explore working in a research team. I would highly recommend other students to participate in the future.