Teaching
I designed the Computer Vision paper discussion section at Indiana University in 2022 and have taught it every spring since, receiving the Luddy School's Associate Instructor of the Year award for the first iteration. I returned in 2026 as a guest lecturer on generative models. The course is led by my advisor, Prof. David Crandall. Over sixteen weeks the section works through the papers that shaped modern visual recognition, from early convolutional networks to vision transformers — the aim being to build intuition and mathematical understanding, not familiarity with a list of names.
Invited to deliver the generative modeling lectures in the same Computer Vision course whose discussion section I designed and ran for four years.
The sequence moved through the main families in the order of how each frames the generation problem: autoencoders and their variational form, where the model learns a compressed latent space it can sample from; GANs, where generation becomes a two-player game between a generator and a discriminator; and diffusion models, which recast generation as learning to reverse a gradual noising process. Taught in that order, the through-line is visible — each is a different answer to the same question of how to model a data distribution well enough to draw new samples from it.
Slides: Week 12 — Autoencoders, VAEs & GANs · Week 13 — Introduction to Diffusion Models
Fourth and most recent iteration, and the fourth consecutive year running the section since designing it in 2022. The sixteen-week structure has proven durable: students still arrive with a fragmented picture of modern architectures and leave able to place any new model in a lineage.
Third iteration. The core sequence — CNNs, MLPs, vision transformers — held up, with the reading list revised to keep pace with the architectures that mattered by then.
Second iteration of the discussion section I designed the previous spring, now settled into a sixteen-week arc across three families of architecture: convolution-based networks, MLP-based networks, and vision transformers.
Running it a second time let me tighten the sequence — which papers genuinely need a full session, which pair well, and where students consistently get stuck. Enrollment grew past 150.
I designed and taught the first iteration of the paper discussion section for Computer Vision — the first of its kind in the department. Rather than a textbook treatment, the section worked directly through the papers that built the field, aiming for both intuition and the mathematics behind each idea.
The reading list traced one long argument: how vision moved from convolution to attention. It opened with the CNN lineage — LeNet, AlexNet, GoogLeNet, the ResNet family and its variants, DenseNet, ConvMixer, Xception — turned to the MLP-based challengers questioning whether convolution was necessary at all (MLP-Mixer, ResMLP, CycleMLP, S²-MLP), and closed on transformers for vision: ViT, DeiT, Swin, LocalViT, CvT, alongside self-attention and shift-based designs.
More than 130 students took part. The format put students in the position of presenting and defending a paper themselves, which is a different skill from following a lecture.