StreamSplit: Adaptive Audio Learning for Edge Devices
StreamSplit is a new framework for continuous audio representation learning on edge devices, addressing the incompatibility of large-batch contrastive learning with resource-constrained environments. It introduces a distribution-based streaming approach with a Hybrid Loss to decouple representation quality from local batch size, enabling practical streaming CL on heterogeneous ARM platforms. The framework targets models like CLAP and COLA, adapting to runtime volatility without static compression.
Key facts
- StreamSplit is a framework for streaming contrastive learning on edge devices.
- It addresses the conflict between continuous audio and discrete batch requirements.
- Uses a distribution-based streaming framework with Hybrid Loss.
- Targets ARM client platforms.
- Works with models like CLAP and COLA.
- Decouples representation quality from local batch size.
- Adapts to runtime volatility of edge environments.
- Published on arXiv with ID 2605.26523.
Entities
Institutions
- arXiv