My code and experiments following along with MIT's deep learning lectures. Each folder corresponds to a lecture and contains the notebooks / scripts I wrote while working through that topic.
| Folder | Topic |
|---|---|
lecture_9 |
Transformers and a ResNet test (transformer.py, transformer.ipynb, test_resnet.ipynb) |
lecture_10 |
LSTM (lstm.ipynb) |
lecture_11 |
Denoising autoencoder that reconstructs clean images from noisy MNIST digits (denoising_autoencoder.ipynb) |
lecture_12 |
Self-supervised learning: SimCLR contrastive training on CIFAR-10, and a CLIP vs. DINOv2 comparison using embeddings + K-NN for classification (simCLR_cifar.ipynb, compare_clip_dinov2.ipynb) |
lecture_13 |
Information bottleneck: reproducing the "information plane" experiment (Shwartz-Ziv & Tishby, 2017) on MNIST to see whether a network compresses during training (mnist_binning.ipynb) |
lecture_14 |
Generative models — see breakdown below |
A survey of the main families of generative models, each in its own notebook.
| Notebook | What it does |
|---|---|
density_estimation.ipynb |
RealNVP normalizing flow on a 2D toy distribution |
pixel_cnn.ipynb |
PixelCNN autoregressive image model |
gan_generator_mnist.ipynb |
Vanilla GAN generating MNIST digits |
dcgan_cifar10.ipynb |
DCGAN on CIFAR-10 |
wgan-gp_cifar10.ipynb |
Wasserstein GAN with gradient penalty on CIFAR-10 |
ddpm_mnist.ipynb |
Denoising diffusion probabilistic model (DDPM) on MNIST |
classifier_free_guidance.ipynb |
Classifier-free guidance for conditional diffusion |
latent_diffusion_model_vqvae.ipynb |
VQ-VAE that encodes images into the latent space for latent diffusion |
latent_diffusion_model_ddpm_class.ipynb |
Class-conditioned latent DDPM (UNet + linear noise scheduler) |
latent_diffusion_model_ddpm_image_cond.ipynb |
Image-conditioned latent DDPM |
Shared building blocks live in lecture_14/blocks/ (ddpm_blocks.py,
vqvae_blocks.py) and lecture_14/models/ (vqvae.py).
This project uses uv and Python 3.13.
uv syncDatasets (MNIST, CIFAR-10, CelebA-HQ) are downloaded into each lecture folder on first run and are git-ignored.