Deep Unsupervised Learning
University of Oxford
First Class (A+)
Six projects in deep unsupervised and generative learning—from autoencoders and VAEs to GANs, diffusion models, and CLIP.
Hands-on implementations spanning representation learning, generative computer vision, and multimodal retrieval: reconstructing corrupted images, class-conditioned generation, paired and unpaired translation, Stable Diffusion fine-tuning, and semantic image search.
Six projects tracing the arc of deep unsupervised and generative learning: reconstruction → latent generation → GAN translation → diffusion → multimodal retrieval.
Representation & Generative Foundations
Clustering, PCA, autoencoders, VAE, GAN intro
- Image Inpainting with Autoencoders
- Conditional VAE
Generative Computer Vision
Conditional GAN, image-to-image translation, unpaired translation
- Pix2Pix Image Translation
- CycleGAN
Diffusion, Multimodal & Applications
Self-supervised learning, diffusion models, representation learning
- Stable Diffusion Fine-tuning
- CLIP Image Search
This course gave me a genuine arc—from fixing broken pixels with autoencoders to steering entire image worlds with text. Each week stacked on the last, and by the end I could feel how modern generative AI is assembled from pieces I had actually touched: latents, noise schedules, discriminators, CLIP embeddings.
What excited me most was watching abstractions turn into pictures. Skip connections really do rescue detail. Cycle consistency really does make unpaired translation possible. Fine-tuning only the UNet really is enough to nudge a whole aesthetic. None of this stayed theory on a slide—it became something I could train, evaluate, and show.
I left Oxford more confident that I can learn hard systems by building them, and more optimistic about where this field is going. Diffusion and multimodal models are not magic black boxes to me anymore—they are pipelines I understand well enough to extend, debug, and dream on top of.
Image Inpainting with Autoencoders
Compare a plain conv autoencoder against a U-Net autoencoder for filling randomly corrupted CIFAR-10 regions.
Conditional VAE
Class-conditioned CIFAR-10 generation with a convolutional β-VAE and one-hot label conditioning.
Pix2Pix Image Translation
Paired edge-to-shoe translation on 49k Edge2Shoes samples with a U-Net generator and PatchGAN discriminator.
CycleGAN
Unpaired selfie ↔ anime translation on 3,400 training images with cycle-consistency and identity losses.
Stable Diffusion Fine-tuning
UNet fine-tuning on 1,221 Naruto-captioned images atop a small Stable Diffusion checkpoint—with full DDPM training and inference pipelines.
CLIP Image Search
Zero-shot classification and top-K text/image retrieval over 10,000 Tiny-ImageNet images with CLIP ViT-B/32.
University of Oxford
Lady Margaret Hall
First Class (A+, 85)
15 CATS / 7.5 ECTS / 4 US Credits
Work completed during Oxford summer coursework. This page shows only my own implementations and outputs; course materials are not redistributed.