Similarity Learning
A technique in AI focusing on training models to measure task-related similarity between data points.
Similar to supervised and unsupervised approaches within the broader field of AI, similarity learning is instrumental for tasks requiring effective comparison of data points, such as in clustering, ranking, or information retrieval systems. It involves designing models that can discern the degree of similarity between inputs through learned representations, often leveraging metric learning approaches to ensure these models effectively generalize across unseen data. Its applications are extensive, ranging from facial recognition to recommendation engines, and it frequently employs neural networks like Siamese networks or triplet networks to optimize for similarity-based objectives. Central to its theory is the idea of transforming data into an embedding space where distances between points correspond to their semantic similarities, thus allowing for intuitive, task-specific comparisons.
The earliest ideas relating to similarity emerged from fields like information retrieval and pattern recognition in the 1960s. However, the concept of similarity learning gained particular popularity in the AI community in the 1990s with the development of new metric learning techniques adapted for more complex, high-dimensional data.
Yann LeCun and Léon Bottou are notable contributors to developing similarity learning methods, particularly with the introduction and exploration of Siamese networks in the 1990s. More recently, researchers like Geoffrey Hinton and teams at Google Research have further advanced the field with innovative neural network architectures designed for optimal representation learning within similarity tasks.