FlashPapers
OpenAI (2021)

Learning Transferable Visual Models From Natural Language Supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever

TL;DR (Too Long; Didn't Read)

Instead of training computer vision models on rigid, pre-defined sets of class labels, CLIP jointly trains an image encoder and a text encoder using a contrastive loss on 400 million internet-sourced image-text pairs. This allows the model to perform highly flexible, zero-shot image classification on downstream tasks simply by writing natural language prompts.

400M Web Dataset (WIT) 76.2% Zero-Shot ImageNet Unprecedented Robustness Prompt Engineering Driven

Why This Matters

For over a decade, computer vision research was constrained by a supervised learning paradigm. Models were trained on datasets like ImageNetA benchmark dataset containing over 14 million annotated images across 1,000 distinct object classes. to predict a fixed set of pre-determined categories.

This approach has two major flaws:

  • High Annotation Costs: Labeling millions of images with exact target categories requires massive crowd-sourced human effort.
  • Extreme Brittleness: Supervised models learn specific dataset features rather than general visual concepts. When tested on slight distribution shifts (e.g., sketches, art, or adversarial examples), their accuracy plummets.

The Big Idea

CLIP shifts the paradigm from predicting discrete labels to predicting natural language pairings. Instead of deciding if an image is class #342, the model evaluates how well a text description matches an image.

By leveraging 400 million image-text pairs collected passively from the internet, CLIP learns visual concepts directly from raw language supervision. This unified representation space binds language and vision together, enabling seamless zero-shot transfer to new tasks.

How It Works

1. Contrastive Pre-training

Given a batch of $N$ (image, text) pairs, the model passes the images through an Image Encoder (such as a ResNet or a Vision Transformer (ViT)An architecture that applies transformer blocks directly to sequences of image patches rather than using standard convolutions.) and the texts through a Text Encoder (a standard Transformer).

These encoders produce embeddings $I_1, I_2, \dots, I_N$ and $T_1, T_2, \dots, T_N$ which are projected into a shared multi-modal embedding space. CLIP's objective is to maximize the cosine similarity of the $N$ correct pairings $(I_i, T_i)$ while minimizing the similarity of the $N^2 - N$ incorrect pairings $(I_i, T_j)$ where $i \neq j$.

Symmetric Cross-Entropy Loss over Cosine Similarities

$$\mathcal{L} = \frac{1}{2} \left( \mathcal{L}_{\text{image}} + \mathcal{L}_{\text{text}} \right)$$

Where individual losses are calculated using softmax scaled by a learnable temperature parameter $\tau$.

Interactive Widget 1: Contrastive Loss Matrix Simulator

Adjust the temperature parameter $\tau$ to see how it scales the logits and sharpens/flattens the resulting probability matrix for a toy $4 \times 4$ batch.

Target diagonal matrix Perfect pairing is highlighted along the main diagonal.
Raw Cosine Similarities ($\cos(I_i, T_j)$)
0.85
0.12
0.05
-0.15
0.08
0.78
0.21
0.02
-0.05
0.15
0.91
0.10
0.11
-0.08
0.03
0.82
Softmax Probabilities ($P(i \text{ matches } j)$)
Notice how lower temperatures ($\tau \to 0$) sharply amplify high similarities, helping the model strongly distinguish correct pairings during pre-training.

2. Zero-Shot Classifier Synthesis

At test time, CLIP bypasses the traditional fixed-class classification layer. Instead, we write natural language descriptions of the target classes. The text encoder embeds these descriptions, creating a set of dynamic weights.

We then embed the query image, calculate its cosine similarity against all class description embeddings, and pass those scores through a softmax to output a probability distribution.

Interactive Widget 2: Zero-Shot Classifier Playground

Select a toy image and input custom class names to simulate how CLIP dynamically generates a classification distribution.

Simulated Softmax Probability
*Note: Heuristics simulate text-to-image semantic matching scores based on keyword overlap and base category correlations.

3. Prompt Engineering & Ensembling

A key discovery of the paper is that using raw class names (e.g., "dog") is suboptimal. This is because single words suffer from polysemy (e.g., "boxer" could mean an athlete or a dog breed) and differ from the full-sentence descriptions seen during pre-training.

By wrapping class names in templates like "A photo of a {label}.", zero-shot performance on ImageNet improved by 1.3%. Ensembling predictions across multiple prompts (e.g., "a photo of a small {label}", "a satellite photo of a {label}") boosted accuracy by an additional 3.5%.

Key Results & Robustness

CLIP matches the zero-shot accuracy of a fully supervised ResNet-50 on ImageNet (76.2%) without using any of its 1.28 million training examples.

More importantly, CLIP displays remarkable robustness to distribution shift. When evaluated on out-of-distribution variations of ImageNet, supervised models break down, while CLIP maintains stable performance.

Interactive Widget 3: Robustness Gap Visualizer

Toggle between a standard supervised ResNet-50 and Zero-Shot CLIP to see how the performance gap narrows across shifted distributions.

ImageNet-V2 Natural replication
64.3%
ImageNet-R Renditions & Art
37.7%
ImageNet-Sketch Black & White Sketches
25.2%
ImageNet-A Adversarial examples
2.7%
Supervised models show a severe performance drop when evaluated on shifted distributions. Click "Zero-Shot CLIP" to see the difference.

Limitations & Social Biases

Despite its impressive zero-shot capabilities, the authors note several key limitations of CLIP:

  • Weak on Complex/Abstract Tasks: CLIP struggles on highly specialized or abstract tasks, such as counting objects, satellite image classification, or estimating distances.
  • Poor Out-of-Distribution Generalization: Truly novel, out-of-distribution domains (like handwritten digits in MNIST) result in near-random performance.
  • Social Biases: Since CLIP is trained on uncurated internet data, it inherits and amplifies societal biases. For example, the model exhibits denigration harms, showing higher misclassification rates of certain demographic groups into crime-related or non-human categories.

Glossary

Contrastive Learning

A machine learning technique where a model learns representations by pulling positive (matching) pairs closer together in embedding space while pushing negative (non-matching) pairs far apart.

Zero-Shot Transfer

The ability of a pre-trained model to generalize and perform classification or other tasks on entirely unseen datasets without receiving any task-specific parameter updates or fine-tuning.

Prompt Engineering

The process of carefully designing, structuring, and formatting text prompts to guide a language or multimodal model toward generating the most accurate and context-aware predictions.

Polysemy

The coexistence of many possible meanings for a single word or phrase (e.g., "crane" meaning either a bird or heavy construction machinery).

Cite this work

@inproceedings{radford2021learning,
  title={Learning Transferable Visual Models From Natural Language Supervision},
  author={Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya},
  booktitle={International Conference on Machine Learning},
  pages={8748--8763},
  year={2021},
  organization={PMLR}
}
Made with Flash Papers — make your own