0:00
/
0:00
Transcript

"Expanding Deep Learning-based Sensing Systems with Multi-Source Knowledge Transfer"

The podcast on this paper is generated with Google's Illuminate.

HaKT, proposed in this paper, efficiently transfers knowledge from multiple source models to expand sensing systems with minimal overhead.

This paper introduces HaKT (Heterogeneity-aware Knowledge Transfer), a framework that efficiently expands deep learning-based sensing systems to new domains while handling limited labeled data and heterogeneity challenges. It selects high-quality source models, fuses their knowledge, and adaptively injects it into target models.

-----

https://arxiv.org/abs/2412.04060

🔍 Original Problem:

Expanding deep learning sensing systems to new domains faces three key challenges: limited labeled data, data heterogeneity across domains, and device heterogeneity constraints. Existing approaches like federated learning and transfer learning fail to address these challenges holistically.

-----

🛠️ Solution in this Paper:

→ HaKT first employs a two-stage model selection protocol to identify high-quality source models efficiently without excessive communication overhead.

→ It uses an attention-based mixer to fuse knowledge from multiple source models by assigning sample-specific weights based on feature similarity.

→ The framework includes a knowledge dictionary to store fused predictions and an adaptive learner to optimize knowledge injection into target models.

→ A low-cost joint training scheme simultaneously updates the mixer, target model, and partially frozen source models to reduce computational overhead.

-----

💡 Key Insights:

→ Knowledge from source models can help mitigate label scarcity in target domains

→ Sample-wise weighting is crucial for resolving knowledge conflicts between source models

→ Selective storage and adaptive injection of fused knowledge improves target model training

-----

📊 Results:

→ Achieves up to 16.5% accuracy improvement over state-of-the-art baselines

→ Reduces communication traffic by up to 39%

→ Demonstrates superior performance across various tasks and modalities

Discussion about this video