This post is about turning a photo of a cat into a photo of a goldfish by changing only one pixel, at least according to resnet50. With Organizers we participated in RCTF during the close race at the end 2022 to be #1 on CTFtime. This literally meant to participate in every high rated CTF and solving every challenge, including the miscy of the misc. The challenge catspy appeared at around 2am in the misc category and the description states:
The lecture Machine Learning for the Natural Sciences promises to focus on applications of machine learning to natural sciences, especially physics and chemistry. However, most of the actual content is repeating machine learning basics, that is already in foundational lectures on machine learning. In the remaining time, a few interesting are presented, but sadly just very shallowly. There is also programming homework that counts for 1/3 of the final grade. This is nice, and I think more courses should do that.
With KITCTF we participated in the bo01lers CTF and finished 6th. There were some quite fun challenges. Including the resnet challenge, which is a machine learning challenge. I hope to see more machine learning challenges in the future. Challenge description: A naive AI startup released a new visual password system based on State-of-the-Art Neural Network technology. Wanting to save on costs they reuse the popular Resnet model to create embeddings which input password images are checked against hoping to leverage the feature extraction capabilities of Resnet.
Answers to self test questions for the lecture “Data Science I” at KIT. If you spot any errors, write me an e-mail or Discord message. Lecture 1: IntroductionGive examples of applications of clustering.Customer groups clustered based on bought productsUnsupervised malware family identificationOutlier DetectionDescribe a scenario from natural sciences, in which classification is useful: What are the attributes/class? How would you try to solve it?Flower family classification: Attributes (features)Color of different partsShape of different partsSize of different partsSolve it by training a multi-class NN with enough high quality training dataExplain the principle of the One Rule classifier.
Self test questions for the lecture “Machine Learning - Foundations and Algorithms” at KIT. Lecture 3: Model SelectionWhy is it a bad idea to evaluate your algorithm on the training set?Evaluating on the training set, rewards overfitting. Overfitting means learning training points by heart, instead of approximating the distribution the training points were drawn from. A trivial algorithm that just stores and queries all training points, has 100 % accuracy on the training set.
Overview # Topics from the lecture. I use mind maps like this before exam preparation to look see what I can remember from during the semester. During learning, I expand on the topics. Later I use color coding of the nodes to visualize what topics I know well and what still need practice. I also made an Anki deck. But sadly the deck contains screenshots of slides that are not licenced under a free licence.