The advent of Foundation Models is transforming machine learning across many modalities (e.g., language, images, videos) with prompt engineering replacing training in many settings. Recent work on tabular data (e.g., TabPFN) hints at a similar opportunity to build Foundation Models for classification for numerical data. In this paper, we go one step further and propose a hypernetwork architecture that we call MotherNet, trained on millions of classification tasks, that, once prompted with a never-seen-before training set generates the weights of a trained ``child'' neural-network. Like other Foundation Models, MotherNet replaces training on specific datasets with in-context learning through a single forward pass. In contrast to existing hypernetworks that were either task-specific or trained for relatively constraint multi-task settings, MotherNet is trained to generate networks to perform multiclass classification on arbitrary tabular datasets without any dataset specific gradient descent. The child network generated by MotherNet using in-context learning outperforms neural networks trained using gradient descent on small datasets, and is competitive with predictions by TabPFN and standard ML methods like Gradient Boosting. Unlike a direct application of transformer models like TabPFN, MotherNet generated networks are highly efficient at inference time. This methodology opens up a new approach to building predictive models on tabular data that is both efficient and robust, without any dataset-specific training.

Ensembles of deep neural networks with diverse architectures significantly improve generalization accuracy. However, training such ensembles requires a large amount of computational resources and time as every network in the ensemble has to be separately trained. In practice, this restricts the number of different deep neural network architectures that can be included within an ensemble. We propose a new approach to address this problem. Our approach captures the structural similarity between members of a neural network ensemble and train it only once. Subsequently, this knowledge is transferred to all members of the ensemble using function-preserving transformations. Then, these ensemble networks converge significantly faster as compared to training from scratch. We show through experiments on CIFAR-10, CIFAR-100, and SVHN data sets that our approach can train large and diverse ensembles of deep neural networks achieving comparable accuracy to existing approaches in a fraction of their training time. In particular, our approach trains an ensemble of $100$ variants of deep neural networks with diverse architectures up to $6 \times$ faster as compared to existing approaches. This improvement in training cost grows linearly with the size of the ensemble.