Computer Vision

A 7-layer convolutional network reads handwritten digits with weight sharing, pooling, and end-to-end training.

The blueprint for every modern CNN; everything that follows is a refinement of these ideas.

Backpropagation, basic image processing

Convolution + pooling + non-linearities is the right inductive bias for images.

A deep CNN with ReLU, dropout, and dual-GPU training crushes ImageNet 2012 by a 10-point margin.

The landmark result that accelerated deep learning adoption in vision — the moment deep learning took over computer vision.

LeNet, basic CUDA awareness

Big data + GPUs + ReLU + dropout was the unlock; the architecture was the easy part.

Stacks 3×3 convolutions to depths of 16 and 19 layers, showing depth itself is a strong prior.

The cleanest demonstration that depth matters and a perennial backbone for transfer learning.

AlexNet, basic CNN math

Small kernels stacked deep beat large kernels stacked shallow.

Introduces the Inception module, mixing parallel kernels of multiple sizes within each block.

A counterpoint to VGG showing that smarter blocks beat brute-force depth.

VGG, computational complexity of conv

Multi-scale features can live inside a single layer.

Adds identity skip connections so the network only has to learn residuals; trains 152-layer nets.

The single most important architectural idea after the convolution itself; appears in every modern net.

VGG, vanishing gradients

If you can't train it deeper, give the gradient a shortcut.

Each layer receives inputs from every preceding layer, encouraging feature reuse.

Useful contrast to ResNet; clarifies what skip connections are actually doing.

ResNet

Feature reuse, not just gradient flow, explains why skip connections help.

Normalises layer activations within each mini-batch, dramatically stabilising deep network training.

Pretty much every CNN since uses BN; understanding its dynamics is foundational.

CNN training basics

Stabilising activation distributions during training is half the battle.

Region proposals fed into a CNN and an SVM produce object detections.

The starting point of the modern detection pipeline.

AlexNet, selective search

You can repurpose a classifier into a detector by feeding it proposals.

Reuses one CNN forward pass for all proposals via a region-of-interest pooling layer.

A clean lesson in how one architectural idea (RoI pooling) can collapse a whole pipeline.

R-CNN

Compute features once, then crop them — don't re-run the CNN per proposal.

Replaces external proposals with a learned region proposal network sharing the detector's features.

The end-to-end detection paper; basis of two-stage detectors for years.

Fast R-CNN

Make every component of the pipeline learnable.

Treats detection as a single regression problem from image to bounding boxes and classes.

The other branch of the detection family — single-stage, fast, deployable.

CNN basics

Sometimes you should just predict the answer directly.

A modified cross-entropy that down-weights easy examples, fixing class imbalance in dense detectors.

A self-contained lesson in loss design; the focal loss now lives in many other domains too.

YOLO, cross-entropy

When most examples are trivial, weight by difficulty.

Replaces the final FC layers of a classification CNN with conv layers to produce per-pixel labels.

The simple, beautiful idea that turned classifiers into segmenters.

VGG, transposed conv

A classifier without dense layers is a segmenter.

A symmetric encoder–decoder with skip connections that fuses low-level detail and high-level context.

Far more than a medical paper — U-Net is the backbone of modern diffusion models too.

FCN

Symmetric encoder–decoder with skips is a remarkably general architecture.

Adds a small fully-conv mask branch to Faster R-CNN, producing instance segmentation alongside detection.

Shows how a tiny architectural addition unlocks a whole new task.

Faster R-CNN

Multi-task heads atop a shared backbone are absurdly effective.

A generator and discriminator play a minimax game; samples come from the generator at convergence.

A foundational generative-modeling idea; understanding GANs sharpens your intuition for diffusion.

Game theory basics, CNNs

You can train a generator without a likelihood by pitting it against a learned critic.

A set of architectural guidelines (strided conv, batch norm, no FC) that finally make GANs train stably.

The first pretty GAN samples and a study in how much architectural detail matters in adversarial training.

GANs

GAN training is mostly a stability problem.

Disentangles style and content via adaptive instance norm and a learned latent W space.

The pinnacle of GAN-era image quality and a beautiful study in disentanglement.

DCGAN

Architectural inductive biases can produce semantically meaningful latent spaces for free.

Splits an image into patches, treats them as tokens, and feeds them into a vanilla Transformer encoder.

The convergence point of NLP and vision — proves convolutions aren't architecturally necessary.

Transformer, ResNet

With enough data, attention beats convolution at images.

Detection as set prediction: a Transformer outputs a fixed set of boxes, matched via Hungarian loss.

A radical re-framing of detection that removed anchors, NMS, and most heuristics.

Transformer, Faster R-CNN

Many "necessary" hand-tuned components turn out to be replaceable by attention + the right loss.

Restricts self-attention to local shifted windows, restoring CNN-like multi-scale efficiency.

The bridge between ViT's purity and a CNN's practical efficiency.

ViT

Locality and hierarchy can be reintroduced into Transformers when compute is tight.

Masks 75% of image patches and reconstructs them with an asymmetric encoder-decoder ViT.

Brought BERT-style masked pretraining to vision — now a default self-supervised recipe.

ViT, BERT

Images are redundant enough that aggressive masking makes a great pretext task.

Modernises a ResNet step-by-step with ViT-era training tricks until it matches Swin Transformer.

A careful ablation showing how much of the ViT gap was training recipe, not architecture.

ResNet, Swin Transformer

Attribute gains carefully — recipes and architectures are easy to conflate.

Trains image and text encoders contrastively on 400M web image-caption pairs.

The bedrock of modern multimodal AI; underlies retrieval, classification, and image-generation guidance.

Transformers, contrastive learning

Natural language is a far more general supervision signal than fixed label sets.

A 12B-parameter Transformer trained autoregressively over discrete image-token + text-token sequences.

The first compelling text-to-image system and a clean conceptual ancestor of all diffusion T2I.

GPT-3, VQ-VAE

You can model images as discrete tokens and generate them with a language model.

Trains a network to reverse a fixed Gaussian noising process, generating images by iterative denoising.

The paper that revived and re-formalised diffusion models. Read this before any T2I paper.

GANs, basic stochastic processes

Reversing noise step-by-step is a simpler, more stable generative recipe than adversarial training.

Runs the diffusion process in a compressed VAE latent space, slashing compute by an order of magnitude.

The paper that put open-source text-to-image into millions of hands.

DDPM, VAE

Don't diffuse pixels — diffuse a small, perceptually-equivalent latent representation.

Pairs a frozen large language model text encoder with a cascaded pixel-space diffusion decoder.

A clean, practical study showing that text-encoder quality dominates image-decoder size.

DDPM, CLIP

A bigger text encoder beats a bigger image decoder.

A promptable segmentation foundation model trained on 1B masks across 11M images.

The "major breakthrough moment" for segmentation — a single model that segments anything from any prompt.

ViT, segmentation basics

Foundation models work in vision too — given the right task and a billion labels.

A self-supervised ViT trained on a curated 142M image dataset producing all-purpose features.

The strongest pure-vision foundation model — a counterpoint to CLIP's language-supervised approach.

ViT, self-supervised learning

You can build vision foundation models without text — careful data and self-supervision suffice.

Encodes a scene as an MLP mapping 5D coordinates to colour and density, rendered by ray marching.

Founded neural 3D representation — a whole subfield traces to this paper.

Volume rendering basics, MLPs

A neural network can be the scene representation itself, not just a processor of one.

Represents scenes as millions of optimised anisotropic 3D Gaussians, rasterised in real time.

Displaced NeRF for many practical uses within a year — a lesson in explicit vs implicit representations.

NeRF, rasterisation basics

An explicit, differentiable primitive can beat an implicit network on both speed and quality.