Quentin Herau’s personal page
Intern March-August 2021
Research topics with IPAL
Several models for collaborations between human and machine learning have been proposed. In the context of reinforcement learning, human feedback is used as a reward value or as policy labels to train an agent. In the context of active learning, the machine chooses data samples for humans to label, which is then fed into the training process either in an online fashion or in batches. In human-machine co-learning , humans and machines iterate through the labeling process to improve the quality of labels.
Unlike learning from a large number of pre-labeled data samples, human inputs in human-in-the-loop learning have a larger impact or even overriding effects on machine decisions. Such human-AI collaboration models made it possible for malicious humans to impact the outcome of machine learning models. These issues extend to semi-supervised techniques and few-shot learning techniques, where only a small number of labeled data are used for training, and thus, maliciously labeled data can have consequences on the output.
In this internship, we plan to build upon existing techniques such as expected model output change (EMOC) to design networks and training strategies that depend on the desirable EMOC. We can then choose how to trade-off between impact of input vs. how fast the model can learn from humans.
Collaborators at IPAL
- Ooi Wei-Tsang
- Carlier Axel
Link to webpage
You can know more about me here: https://www.linkedin.com/in/quentin-herau-38378b140/