Project Vista


Deep convolutional neural networks (CNNs) have revolutionized computer vision, achieving near-human accuracy on many prediction tasks. This has led to a growing interest in using deep CNNs to exploit images, video, and other unstructured data sources for analytics in many applications. However, there are still numerous practical bottlenecks in using deep CNNs for data analytics, including for transfer learning with deep CNNs and interpeting such models in new analytics contexts.

In this project, we build a new data system to enable seamless and efficient exploitation of deep CNNs for emerging multimodal analytics workloads. Rather than segregating multimedia data away in siloed systems, Vista will enable popular and scalable data analytics systems to truly see such forms of data.

Vista currently supports large-scale transfer learning from deep CNNs for reliable, efficient, and easy-to-use multimodal analytics at scale in the Spark-TensorFlow environment. Please see the paper below for more details.

Downloads (Paper, Code, Data, etc.)

  • Materialization Trade-offs for Feature Transfer from Deep CNNs for Multimodal Data Analytics
    Supun Nakandala and Arun Kumar
    SysML 2018 Short paper/poster | Paper PDF

Student Contact

Supun Nakandala: snakanda [at] eng [dot] ucsd [dot] edu


This project was/is supported in part by a Hellman Fellowship.