Information-theoretical label embeddings for large-scale image classification

By François Chollet. 2016

17 citations recorded by the source corpus.

We present a method for training multi-label, massively multi-class image classification models, that is faster and more accurate than supervision via a sigmoid cross-entropy loss (logistic regression). Our method consists in embedding high-dimensional sparse labels onto a lower-dimensional dense sphere of unit-normed vectors, and treating the classification problem as a cosine proximity regression problem on this sphere. We test our method on a dataset of 300 million high-resolution images with 17,000 labels, where it yields considerably faster convergence, as well as a 7% higher mean average precision compared to logistic regression.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Information Theory

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Information Theory

Citations in the corpus

Unified Contrastive Learning in Image-Text-Label Space (2022, methodology): “In this sense, our work is also related to conventional zero-shot classification [9, 28, 45, 62, 65, 66].”
Separating Self-Expression and Visual Content in Hashtag Supervision (2017, methodology): “Second, we follow [3] and learn the parameters θ of the image model f ( · ; θ ) by minimizing the negative cosine similarity between the image embedding, f j = f ( I j ; θ ) ∈ R D , and the sum of the embeddings of the hashtags, h j , corresponding to image I j : A potential advantage of this…”
Large-Scale Visual Relationship Understanding (2018, unknown): “Chollet [3] exploited multi-label annotations, i.”