The Story of My Life
Weeks pass by & work towards my dissertation doesn’t seem to end. Some problems exist because we are privileged. Think how many people have opportunities to pursue PhD & live in a part of the world where technology is so advanced that you have a chance to learn & explore science first hand.
I mean that’s how I was raised. To always be grateful for the resources you have & use resources to help others in any way possible. Once in a lifetime, you get to a point where something matters more than yourself. For me, doing my PhD was one of them. Of course, I am grateful for it & also pissed sometimes. 😁
Anywho 🤷🏻♀️
So this past week I made good progress in drafting research questions and finding gaps in literature. Once I was confident enough by Capricorn standards (LMAO), I reached out to my research committee & gave them a pre-dissertation presentation on a problem I wanted to solve.
They were legit impressed (😉) because Deepfake is so new to the research world. Very few researchers are focused on Deepfake detection & its impact around us. I was super thrilled when I got to know that govt. agencies are interested in Deepfake research & encouraging people to find solutions for it.
For the second phase of my research, I decided to pre-train the models EfficientNet-B4 & B7, then compare them against a benchmark deepfake dataset. Which is awesome, as if we plug those models into the real world & it learns on its own by encountering Deepfakes on a regular basis & detects them as an outcome.
yeah yeah pictorial representation is a Key 🤭
A quite simple yet effective way of comparing two state-of-the-art EfficientNets B4 & B7.
My plan for the coming week is to train & collect performance data for both phases of research & compare them with McNemar’s chi-square using contingency tables to understand true positives, false positives, true negatives & false negatives for EfficientNet-B4 & B7.
McNemar’s chi-square is a suitable statistical test for machine learning models and more convenient for a classification problem like deepfake detection; thus, detection techniques results are binary (e.g., an image or video content is deepfake or actual).
BRAVO BRAVO 👏 until next time!!
well well, don’t forget to skim through wisdom!!! by Jasmin Bharadiya
