NLT

NLT models parameter-level domain shifts between source and target neural networks to adapt crowd counting models across different datasets.

Key Features:

• Domain-Invariant Feature Extraction: Contrasts with CDCC methods that use image translation and adversarial learning by emphasizing parameter-level shifts rather than only image- or feature-space invariance.
• Parameter-Level Domain Gap Description: Represents domain differences via a neuron linear transformation that uses domain factors and bias weights to describe shifts in model parameters.
• Learning from Limited Data: Learns per-neuron parameters from few labeled target-domain samples in a pre-trained source model to enable generation of corresponding target-domain neurons by linear transformation.
• Robustness and Effectiveness: Demonstrates superior performance to alternative domain adaptation, supervised learning, and fine-tuning approaches across six real-world datasets and validated via ablation study.

Scientific Applications:

• Crowd Counting: Enhances adaptability and accuracy of crowd counting models when deployed across different environments or datasets, supporting applications such as public safety management.

Methodology:

Describe parameter-level domain gaps by comparing model parameters between source and target domains and implement a neuron linear transformation that uses learned domain factors and bias weights; learn per-neuron transformation parameters from few labeled target samples in a pre-trained source model and generate target-domain neurons via linear transformation.