**Deep convolutional neural networks as a tool for vision science** **Astrid Zeman**<br/> *Laboratory of Biological Psychology, Department of Brain and Cognition, KU Leuven, Belgium* **Martin Hebart**<br/> *Laboratory of Brain and Cognition, National Institutes of Health, Bethesda MD, USA* **José Oramas**<br/> *ESAT-PSI, IMEC, KU Leuven, Belgium* Together we will cover the fundamentals of deep Convolutional Neural Networks (CNNs), and delve into the ways that we can represent, and interpret, these networks. CNNs are currently the most successful machine learning technique for image classification, becoming immensely favourable among computer vision scientists. CNNs are also a popular model among visual neuroscientists and perceptual scientists due to their significant representational similarity to brains (from single-cell recordings in monkeys to human fMRI and MEG), as well as behaviour. Following on from the main lecture content, we will have a panel discussion on questions that include: Where is the field heading in the future? Where are there big gaps in performance and biological plausibility? What are the different perspectives of computational versus perceptual scientists? What is the big aim that we are all pursuing? We hope to engage everyone to participate in a lively debate. This tutorial will cover three main themes: 1. Fundamental concepts<br/> a. Operations (convolutions, pooling and ReLU) and inference<br/> b. Network architectures and the types of layer connectivity<br/> c. Backpropagation and its goal<br/> d. Seminal papers 2. Representational Similarity Analysis (RSA)<br/> a. Quantifying the relationship between CNNs, brain and behavior<br/> b. Variance partitioning for disentangling unique and shared contributions of different models<br/> c. Advanced methods that allow us to improve the fit in the relationship between brain and behaviour 3. Model interpretation<br/> a. Methods aimed at understanding what a given CNN has actually learned<br/> b. Explaining models and justifying the decisions they make<br/> c. Protocol for evaluating these different techniques