Set2Model Networks: Learning Discriminatively To Learn Generative Models

TL;DR

This paper introduces Set2Model networks, a meta-learning approach that quickly learns generative models from small datasets, improving web-based image retrieval by enabling discriminative training and effective concept modeling.

Contribution

The paper presents a novel end-to-end differentiable meta-learning framework for training Set2Model networks that map example sets to generative models, with a new backpropagation method through mixture fitting.

Findings

Set2Model networks outperform strong baselines in web image retrieval tasks.

The approach effectively models polysemous and noisy concepts.

The method enables rapid concept learning from limited data.

Abstract

We present a new "learning-to-learn"-type approach that enables rapid learning of concepts from small-to-medium sized training sets and is primarily designed for web-initialized image retrieval. At the core of our approach is a deep architecture (a Set2Model network) that maps sets of examples to simple generative probabilistic models such as Gaussians or mixtures of Gaussians in the space of high-dimensional descriptors. The parameters of the embedding into the descriptor space are trained in the end-to-end fashion in the meta-learning stage using a set of training learning problems. The main technical novelty of our approach is the derivation of the backprop process through the mixture model fitting, which makes the likelihood of the resulting models differentiable with respect to the positions of the input descriptors. While the meta-learning process for a Set2Model network is discriminative, a trained Set2Model network performs generative learning of generative models in the descriptor space, which facilitates learning in the cases when no negative examples are available, and whenever the concept being learned is polysemous or represented by noisy training sets. Among other experiments, we demonstrate that these properties allow Set2Model networks to pick visual concepts from the raw outputs of Internet image search engines better than a set of strong baselines.