Research line leader: Max Welling (UvA)
R3 challenges are:
- How to make deep learning predictions more transparent through visual and textual evidence?
- Can we make deep learning predictions more reliable through better confidence intervals?
- Can we make the design of neural networks more efficient by providing an interactive deep learning dashboard?
Our current understanding of deep learning is lagging behind its immense success, hence the reason why DL algorithms are often referred to as ‘black box methods’. In order to continue its success story and to make it a reliable tool set for practical applications, deep learning needs to be able to account for its results. Especially when it comes to high-risk applications such as medical diagnosis, it is crucial to understand the decision-making process and have well calibrated confidence intervals.
A promising direction towards more transparent deep neural networks (DNNs) aims at providing visual evidence from input data in favor or against decision outputs of deep networks. This allows one to gain a deeper understanding of a network’s inner workings, which in turn provides opportunities to improve efficiency. Moreover, visualizations enable end-users to assess the quality of predictions and combine this with expert knowledge to take more reliable decisions. In this research line, verifiability and accountability of neural networks will be improved by building upon previous work on visualization and confidence estimation.
In parallel, the reliability of deep learning methods will be improved with interactive network training. High accuracy results with DNNs are currently achieved by investing many hours in optimizing the network layout, optimization methods, loss functions, and so forth. However, their joint effect is poorly understood theoretically, condemning data scientists to trial-and-error. Using a deep learning dashboard, this design process will be supported during training by visual analytics. This allows for more efficient interactive network training, which will significantly increase the insight in the task at hand and lead to substantial shorter trail-and-error loops.