# Zero-Shot Learning Through Cross-Modal Transfer

**Authors:** Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani, Christopher D. Manning, Andrew Y. Ng  
**arXiv:** [1301.3666](https://arxiv.org/abs/1301.3666)  
**Date:** 2013-01-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1301.3666v1/

## Key points

- The work presents a novel model for object recognition in images without training data
- The model uses unsupervised large text corpora to acquire knowledge about unseen categories
- Distributional information in language serves as a semantic basis for understanding the appearance of objects
- The proposed model achieves state-of-the-art performance on classes with thousands of training images and reasonable performance on unseen classes
- Two key steps are used: outlier detection in the semantic space and the use of two separate recognition models
- The model learns to recognize objects solely based on textual information from large corpora, without relying on manually defined semantic features for words or images
- This approach bridges the gap between language and visual perception, opening up new possibilities for object recognition without extensive labeled training data.

## Summary

This work, titled "Zero-Shot Learning Through Cross-Modal Transfer," presents a novel model for object recognition in images even when no training data is available for the objects. The model leverages unsupervised large text corpora to acquire knowledge about unseen categories. In this zero-shot framework, distributional information in language serves as a semantic basis for understanding the appearance of objects. Unlike most previous zero-shot learning models that can only differentiate between unseen classes, the proposed model achieves state-of-the-art performance on classes with thousands of training images and reasonable performance on unseen classes. This is accomplished through two key steps: outlier detection in the semantic space and the use of two separate recognition models. Importantly, the model does not rely on manually defined semantic features for words or images; instead it learns to recognize objects solely based on textual information from large corpora. By bridging the gap between language and visual perception, this approach opens up new possibilities for object recognition without extensive labeled training data. The authors of this work are Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani, Christopher D. Manning and Andrew Y. Ng whose research contributes to advancing the field of zero-shot learning and has implications for various applications in computer vision and artificial intelligence.