The world, and particularly New Zealand, is experiencing a shortage of deep learning specialists, and it is unlikely that this will change soon: recent developments have shown that our modern, data-rich world consistently provides new opportunities for machine learning to increase productivity and yield better decision-making. Potential applications proliferate, for example, in the agritech sector, and need to be tackled with a high degree of immediacy so that industry remains competitive. Access to deep learning technology for industry and research institutions wishing to develop their own applications is compromised by the need for expert knowledge in deep learning architectures and algorithms, and the difficulty of developing deep learning models for new domains outside standard application areas such as object recognition, text classification, and speech recognition. Currently, for organisations to successfully apply deep learning to their data, substantial expertise in deep learning techniques must be available in addition to domain expertise required to label examples for machine learning. Moreover, because domain experts are not directly involved in applying machine learning algorithms, opportunities for reducing the number of expert-provided labels by more directly guiding the learning algorithm may be lost. This is a problem because an expert’s time is often very valuable.
The goal of this project is to enable domain experts to apply deep learning without involving a machine learning expert and without requiring any programming, while minimising the amount of data labeling required. Our hypothesis is that a carefully designed software platform that engages the end-user in the deep learning process through an interactive graphical user interface (GUI), automates model selection and parameter tuning, and does not require any programming, will enable access to deep learning technology for a much wider sector of the economy. Moreover, enabling end-users to build predictive models directly without involving deep learning experts will yield more accurate solutions in less time. Our software will be developed by building on work in human-computer interaction, model visualisation, meta learning, active learning, semi-supervised learning, and automatic model selection. We propose a two-phase process. In the first phase, semi-supervised model selection based on meta learning will be applied to choose and adapt a suitable model for the target domain from a library of models garnered from related domains (a model zoo). A small amount of initial expert-labeled data will be required in this phase. In the second phase, semi-supervised active learning will be applied to enable the expert to interactively improve the accuracy of this initial model. To achieve this algorithmically, semi-supervised model selection and robust semi-supervised active learning are the main research problems to be tackled.