PAPERS-02-001 · Paper — The Lineage — How Machines Learned to See

The paper

This is the track opener, so there is no single paper behind it — it is the map you read before the journey. Every other entry in PAPERS-02 teaches one landmark paper in computer vision's lineage; this video walks the whole arc as one story. In roughly three decades the field went from hand-tuned edge detectors to networks that recognize ten thousand object categories, generate photorealistic images, and read a scene the way a language model reads a sentence. The chain is tight: each breakthrough answers the exact bottleneck of the one before it.

Before this paper

Before deep learning, computer vision was feature engineering. A human decided what mattered — edges, corners, gradients, textures — and wrote detectors by hand (SIFT, HOG, Haar cascades). A classifier sat on top of those features, but the features themselves never learned. Progress meant a cleverer hand-designed descriptor. Recognition was brittle: change the lighting, the pose, or the background and accuracy collapsed. The field needed features that were learned from data, not carved by an engineer — and a dataset and a machine big enough to learn them.

The key idea

The arc is a chain of unlocks. LeNet (1998) showed a convolutional network could read handwritten digits end to end — learned filters, not hand-built ones. AlexNet (2012) scaled that idea to natural images and, powered by GPUs and ImageNet, cut error nearly in half overnight — the moment deep learning "won" vision. VGG and GoogLeNet (2014) made networks deeper and more efficient. ResNet (2015) solved the degradation that stopped networks from getting deeper, training past 100 layers with residual connections. GANs (2014) and U-Net (2015) taught networks to generate and to segment. Vision Transformers (2020) dropped convolution entirely, treating an image as a sequence of patches, and CLIP (2021) tied vision to language so a model could recognize things it was never explicitly trained on.

How it works

Read it as a sequence of removed limitations. Convolution gives translation-invariant, weight-shared feature detectors — the same edge filter slides across the whole image, so the network learns far fewer parameters than a dense net and generalizes across position. Stacking convolutions builds a hierarchy: edges → textures → parts → objects. Depth is the lever for richer hierarchies, but naively deep nets stopped improving — until residual connections let gradients skip layers and made depth free. Then attention arrived from language: a Vision Transformer cuts the image into patches, embeds each like a token, and lets every patch attend to every other, learning global relationships a small convolution never could. Finally, training on image–text pairs replaced fixed label sets with open-vocabulary understanding.

See it — the road from edge detectors to open-vocabulary vision

timeline
    title How machines learned to see
    1998 : LeNet convolutions read digits
    2012 : AlexNet GPUs plus ImageNet win
    2014 : VGG GoogLeNet GANs deeper and generative
    2015 : ResNet residual connections go very deep
    2020 : Vision Transformer image as patches
    2021 : CLIP vision tied to language

What they showed

Because this opener is a narrative, it reports no numbers of its own — each downstream video carries its paper's actual headline result, cited to the arXiv id or original venue. What the arc as a whole showed is decisive and measurable: learned convolutional features beat every hand-engineered descriptor on ImageNet; residual connections turned "deeper hurts" into "deeper helps," winning ILSVRC 2015; and a transformer with enough data matched or beat the best convolutional nets while transferring better. Each result answered a specific doubt about the last idea, which is why the lineage reads as a single argument rather than a list.

What it changed

Together these ideas turned vision from a craft of hand-tuned detectors into a discipline of learned, general representations — the shift that produced face unlock, medical-image triage, self-driving perception, and text-to-image generation. And the endpoint rhymes with language: an image is now often processed the same way a sentence is, by attention over tokens, which is why a single architecture can increasingly handle text and pixels together. None of the ideas erased the last; the modern vision stack is convolution's inductive bias, residual depth, and attention's global reach, all at once. Next: Gradient-Based Learning (LeNet).

The one-line takeaway

Machines learned to see by learning their own features — convolution, then depth, then attention — each step removing the exact limit of the one before.

Read it / meet it

There is no arXiv id for the opener — it is the through-line, and each linked paper carries its own citation. You meet this lineage every time your phone groups photos by face, a car brakes for a pedestrian, or a model turns a prompt into a picture. Start the track at the next video and follow it root to frontier; by the end you will have read modern computer vision as one continuous idea.