Limited training data

Limited training data

Visual Intelligence aims to develop new deep learning models that solve problems involving complex images from limited training data.

Motivation

The performance of deep learning methods steadily improves with more training data. However, the availability of suitable training data is often limited. Additionally, labelling complex image data requires domain experts and is both costly and time-consuming.

This research challenge is heavily stressed by a majority of our user partners as an immediate need. To succeed in our innovation areas, it is absolutely necessary to research new methodology which learn from limited and complex training data.

Solving research challenges through new deep learning methodology

Methods which exploit weak, noisy and incompletely labelled data, be it through semi-supervised or semi-supervised approaches, make up a significant portion of our portfolio. Examples include the following:

• A self-supervised approach for content-based image retrieval of CT liver images.

• Explainable marine image analysis methods validated on multiple marine datasets, such as multi-frequency echosounder data and aerial imagery of sea mammals captured by drones.

• A self-supervised method for automatically detecting and classifying microfossils.

• Methods for automatic building change detection in aerial images based on self-supervised learning.

These methods represent time-effective and cost-effective approaches which make deep learning models less reliant on large data samples and labeled data. These improve the models’ efficiency and ability to generalize, making them more applicable in real-world settings.

‍