DeepLearning 2.0: Meta-Learning Qualitatively New Components

Introduction

Deep learning has revolutionized many fields, such as computer vision, speech recognition, natural language processing, and reinforcement learning. This success is based on replacing domain-specific hand-crafted features with features that are learned for the particular task at hand.

Objective

The logical step to take deep learning to the next level is to also (meta-)learn other hand-crafted elements of the deep learning pipeline. We therefore propose to develop meta-level learning methods for the creation of novel customized deep learning pipelines, by means of:

1. Hierarchical neural architecture search for learning qualitatively new architectures and architectural building blocks from scratch.
2. Learning of optimizers and hyperparameter adaptation policies that adapt to their context in order to converge faster and more robustly.
3. Learning the data to train on, to remove the need for large sets of labelled data.
4. Bootstrapping from prior design efforts to increase efficiency and make an integrative design of architectures, optimizers, hyperparameter adaptation policies, and pretraining tasks feasible in practice.

Expected Advances

These advances will allow the next generation of deep learning pipelines to achieve higher accuracy, lower training time, and improved ease-of-use (democratization of deep learning). They will also allow customization to particular design contexts, including additional objectives next to accuracy (such as robustness, memory requirements, energy consumption, latency, interpretability, training cost, uncertainty estimation, and algorithmic fairness) in order to facilitate trustworthy AI.

Implementation Plan

In order to demonstrate the effectiveness of these methods, we plan to develop:

5. New state-of-the-art customized deep learning pipelines for various applications, including EEG decoding, RNA folding, and improving the reinforcement learning pipeline and deep learning on tabular data.