Learning to generate classifiers
This work addresses the challenge of few-shot learning for practitioners needing efficient classifiers with minimal data, though it is incremental as it builds on existing attentional and meta-learning techniques.
The authors tackled the problem of generating classification policies for low-data regimes by training a network to produce classifiers conditioned on entire training sets, achieving significantly better performance than several baseline methods on datasets with 10-50 datapoints.
We train a network to generate mappings between training sets and classification policies (a 'classifier generator') by conditioning on the entire training set via an attentional mechanism. The network is directly optimized for test set performance on an training set of related tasks, which is then transferred to unseen 'test' tasks. We use this to optimize for performance in the low-data and unsupervised learning regimes, and obtain significantly better performance in the 10-50 datapoint regime than support vector classifiers, random forests, XGBoost, and k-nearest neighbors on a range of small datasets.