How do you train a self-driving car to make an emergency stop when a pedestrian dashes across the road? Training data used to prime the vehicle’s machine learning (ML) algorithms might include video footage of a similar scenario. However, with infinite variables from weather conditions to traffic behaviour, it’s impossible to cover all bases during machine training.
To perform well in these dynamic real-world situations, ML models require an abundance of high-quality training data, which experts say isn’t always easy to acquire. “Real-world data collection can be difficult, time-consuming or expensive; and we cannot capture all variations in the training dataset,” explained Wenyu Zhang, a Research Scientist at A*STAR’s Institute for Infocomm Research (I2R).
The difference between situations captured in training data and all possible real-world scenarios is known as a domain shift, which is addressed using source-free domain adaptation (DA) to help algorithms adapt in unfamiliar circumstances. Traditional DA either requires impractical amounts of training data or adopts potentially unreliable learn-on-the-fly approaches—neither are ideal for applications with large domain shifts.
Zhang and colleagues proposed a more flexible, less data-heavy solution leveraging batch normalisation: a technique conventionally used to improve the performance of deep learning neural networks by reducing overfitting, increasing the learning rates and improving the generalisation capabilities.
The team implemented small changes to the batch normalisation layer of the neural network architecture which helps algorithm adapt to real-world information rather than relying primarily on training data. Their few-shot DA framework optimises feature normalisation statistics in pre-trained models with a small target domain support set—a method the team found to be accurate and reliable across multiple datasets.
As a proof of concept, the researchers tested their DA framework using image data that exhibited various types of domain shifts ranging from different art styles to pictures of objects photographed in different environments. They found that their approach could adapt to various domain shifts with ease and perform without extensive real-world training.
“We demonstrated its effectiveness in image classification and segmentation across several publicly available datasets,” said Zhang adding that the team is aiming to expand the DA framework for use with other network and data types.