Debuggable deep learning

Deep learning is often called a black box, so how do you diagnose and fix problems in a deep neural network (DNN)? Avesh Singh and Kevin Wu explain how they systematically debugged DeepHeart, a DNN that detects cardiovascular disease from heart rate data. You'll leave with an arsenal of tools for debugging DNNs, including Jacobian analysis, TensorBoard, and DNN unit tests.

Although deep neural networks have shown high accuracy in fields like computer vision, natural language processing, and medicine, they often behave like black boxes. Avesh Singh and Kevin Wu explain the intuitions that went into building DeepHeart, a DNN that detects cardiovascular disease from heart rate data. Next, they explain techniques to debug DNNs, including visualizing activations and training on synthetic data. Avesh and Kevin also analyze gradients for recurrent neural networks, looking for ways to quantify the amount of information loss a network may be experiencing over time. You’ll leave this talk with a greater understanding of how to design and debug a deep neural network for your machine learning task.