• Declarative Transfer Learning from Deep CNNs at Scale - Supun Nakandala


  • Abstract:
    Deep convolutional neural networks (CNNs) achieve near-human accuracy on many image understanding tasks. Thus they are now increasingly used to integrate images with structured data for multimodal analytics applications. Since training deep CNNs from scratch is expensive, transfer learning has become popular: using a pre-trained CNN, one “reads off” a certain layer of CNN features to represent images and combines them with other features for a downstream ML task. Since no single layer will always offer the best accuracy in general, such feature transfer requires comparing many layers. The current dominant approach to this process on top of scalable analytics systems such as Spark using deep learning toolkits such as TensorFlow is fraught with inefficiency due to redundant CNN inference and the potential for system crashes due to mismanaged memory. We present Vista, the first data system to mitigate such issues by elevating the feature transfer workload to a declarative level and formalizing the data model of CNN inference. Vista enables automated optimization of feature materialization trade-offs, distributed memory management, and system configuration. Real-world experiments show that apart from enabling seamless feature transfer, Vista substantially improves system reliability and reduces runtimes by up to 90%.