Image registration is a standard technique in computer vision, defined in the following ways: “directly overlaying one image on top of another” (Reddy et al), “aligning two or more images” (De Vos et al), and the “process of transforming different image datasets into one coordinate system with matched imaging contents”
Dear Reader: I came across a paper I had started drafting in 2019 on Multimodal Machine Learning - at the time a very new and fascinating field that I was considering to apply towards my Ph.D. Lo and behold, here we are nearly six years later and my dissertation
In natural language processing (NLP) researchers use one type of data which is text, in order to perform tasks such as text classification, question-answering, or language generation. Whereas, in computer vision algorithms like support vector machines (SVM) or convolutional neural networks (CNNs) use image data extracted from static images or
Dear Reader: I recently found a trove of papers I wrote back in 2019 and thought to post a few snippets! Convolutional Neural Networks (CNN) gained popularity in 2012 when AlexNet developed by Alexander Krizhevsky, Sustkever, and Hinton won that year’s ILSVRC (ImageNet Large-Scale Visual Recognition Challenge) classifying images
Clearly, we see this MMAI everywhere nowadays. Enter a prompt in Midjourney and you now get an image. From VQGAN-CLIP who had released her code several years ago to offer the first example of using text prompts to generate images, to, RunwayGen that makes videos (like this 4-sec clip I
Once I deciphered how a neural network operates and was able to diagram and code it by hand using nothing by numpy, I was hooked. It was as if I finally broke through and could interpret all the complex diagrams I saw on the internet. Shortly after I came to
Statistics is the tried and true way of evaluating the performance of a machine learning model. I wrote the material below about two years ago when I was in the thick of my PhD. It's a basic overview of many terms you'll hear over and over
I read an excellent paper this morning by Demir from Intel and Ciftci from Binghamtom University entitled How do Deepfakes Move? Motion Magnification for Deepfake Source Detection, published at CVPR. They also published FakeCatcher: Detection of Synthetic Portrait Videos using Biological Signals. The paper posits that by analyzing features of
I don't have to tell you that everything we see online nowadays is increasingly suspicious. In our age of Generative AI and jaw dropping Tik Tok filters, it's hard to know what's real or what's fake. Since my dissertation was focused on
AI with Catherine Ordun
I'm getting back into it after a holiday break and one of the most fascinating areas of AI research is Language Agents. I'm more of a computer vision person, so LLMs are not my bread and butter. But, I've grown increasingly interested in the
Before I defend, I wanted to share my Top 5 Lessons Learned as a PhD Student. If you're thinking about applying for a PhD - especially in computer science (or adjacent) fields - these might pertain to you. However, I think they're relevant for most PhD experiences.
I remember when pjreddie released You-Only-Look-Once (YOLO) several years ago, and installing darknet (https://github.com/pjreddie/darknet). I also laughed when I read his hilarious licensing agreement LICENSE.fuck (https://github.com/pjreddie/darknet/blob/master/LICENSE.fuck). Well, Yolo continues to give back. And now in 2022 we
Consider that you have a pair of images you want to train. In my case, it's a GAN, where I'm improving upon older work from 2020 in the facial-visible-to-thermal GAN [https://github.com/nudro/favtgan]. My paired image consists of a visible and thermal face, simultaneously
I'm embarrassed to say that even at this point in my life, I still have to look up this formatting to crop an image. Once and for all, I'll post it here so that I don't forget it and also in case you'
You'll eventually face a moment where you're going to have to read through an (at-first) complicated set of AI code. You might need to leverage another expert's repo in order to build an experiment. You might be in a company, and found a great
This semester I'm taking a course at UMBC called Causality in Machine Learning by Dr. Osman Gani [https://osmanii.github.io/]. So far, we've studied Simpson's Paradox, Causality and the Potential Outcomes Framework, Average Treatment Effect, Ignorability, Exchangability, Identifiability, Consistency, Structural Causal Models (i.
Below is an excerpt from our paper Ordun, Catherine, Edward Raff, and Sanjay Purushotham. "The Use of AI for Thermal Emotion Recognition: A Review of Problems and Limitations in Standard Design and Data [https://arxiv.org/pdf/2009.10589.pdf]." arXiv preprint arXiv:2009.10589 (2020) for AAAI
Introduced by Goodfellow, et al. (Goodfellow) in 2014, a GAN consists of two functions, a generator and a discriminator, which are usually neural networks but can be any differentiable function. GANs are used to map data from one space to another; typically from a generated distribution \(p_g\) to an
Decision trees are one of the most fundamental algorithms in machine learning. Let's review how to generate one from scratch. Consider the following table about three different attributes about the ripeness of a fruit. We have density, size, and firmness. Our target is "Ripe" or "
> A paper review of Yosinski, Jason, et al. "How transferable are features in deep neural networks?." Advances in neural information processing systems. 2014. [https://proceedings.neurips.cc/paper/2014/file/375c71349b295fbe2dcdca9206f20a06-Paper.pdf] Slides available here [https://docs.google.com/presentation/d/1uZuEPV_Y3apUpZ6DGUkHwpW9m5RwF2QOfqYpKxnzRFk/edit?usp=sharing] . Summary
It's been a very long time since I've last posted, and at last I've come up for some air after writing a few papers and a lot of code. One thing I've come to appreciate in my PhD studies is the frequent
