New theory identifies how neural networks lose learning flexibility in video streaming systems

A research team has identified a fundamental problem affecting how neural networks adapt to changing network conditions in real-world video streaming: they progressively lose their capacity to learn and adjust, a phenomenon known as plasticity loss.

The core problem#

Adaptive video streaming systems use machine learning to optimize viewing quality by automatically adjusting video bitrate based on available bandwidth and user needs. However, when networks behave differently than the conditions experienced during training, current systems struggle to adapt effectively. The researchers found that neural networks suffer from a specific degradation—their ability to remain flexible and learn from new situations diminishes over time, making them less responsive to novel network patterns.

Rethinking neural plasticity#

The team developed a new theoretical framework called Silent Neuron theory to better explain how and why this plasticity loss occurs. They argue that existing methods for tracking "dormant neurons" (neurons that stop contributing to learning) fail to capture the full picture of what causes adaptation failure. Through analysis of how neural signals propagate forward and backward through networks, they demonstrated that a more nuanced understanding is needed to identify which neurons are genuinely limiting the system's flexibility.

The proposed solution#

Based on their theory, the researchers introduced Reset Silent Neuron (ReSiN), a technique that strategically reactivates neurons that have become unresponsive. Unlike passive monitoring, ReSiN actively preserves plasticity by resetting neurons based on information from both forward computations and backward learning signals. The method includes theoretical guarantees about its performance even when network conditions shift unexpectedly.

Real-world results#

Testing on adaptive video streaming systems showed substantial improvements. ReSiN achieved up to 168% higher bitrate selection and 108% better quality of experience (QoE) compared to existing approaches while keeping video smoothness comparable. The technique also performed consistently well in stable network environments, suggesting it works reliably across different scenarios rather than being a specialized fix for one use case.