Cloud-Aided Computer Vision for Autonomous Devices
Main content start
Principal Investigators:
Keith Winstein and Silvio Savarese
TRI Liaison:
Project Summary
We plan to research compression techniques and other tools to enable systems to efficiently split the training or evaluation of a DNN between a device and the cloud. This will allow deployed systems to improve over time, by sharing “noteworthy” events with the cloud and receiving model updates in return.
Cloud-aided computer-vision systems will allow some features of autonomous vehicles to improve over time and be implemented with less cost, less mass (for computing hardware), and less energy consumption. This can accelerate the deployment of SAE Level 3 and Level 4 autonomy.
Research Goals
- New vector-quantization and entropy-coding schemes for neural-network weights and activations.
- Improvement of field-deployed DNNs over time.
- Differential encoding schemes for model updates, compressing relative to models already deployed.
- “Smoothed” quantizers and encoders to allow backpropagation through coding/decoding layers.
- Real-time enhancement of fail-safe driving systems when network communication is available.
- More-efficient use of GPU memory during training.
- Outsourcing low-duty-cycle tasks to shared resources.