Stronger inductive biases for deep learning
Chief investigator
External collaborators
University of Edinburgh - Miguel Jaques, Kartic Subr, Yordan Hristov, Subramanian Ramamoorthy, Daniel Angelov, Todor Davchev, Tim Hospedales, Martin Asenov
Related links
Research topic Perception and Learning, Modelling and Control
Industry application Medical and Surgical, Assistive Robotics, Autonomous Transport, Remote Field Inspection, Manufacturing, Construction, Service robots
Standard architectures for neural networks have numerous problems with interpretability, flexibility and generalisation. This is in large part due to a lack of inductive biases in models and architectures.
This research aims to explore approaches to embed domain knowledge and structural constraints in deep learning models, in order to improve representation learning for a range of tasks and address these challenges.
As part of an extensive, ongoing collaboration with the University of Edinburgh, we have explored approaches to include inductive biases in neural networks, ranging from differentiable simulators, deep probabilistic programming using controller or reward structure, to the use known physical dynamics constraints to guide learning.
Publications
- NewtonianVAE: Proportional Control and Goal Identification from Pixels via Physical Latent Spaces
- Action sequencing using visual permutations
- Learning Structured Representations of Spatial and Interactive Dynamics for Trajectory Prediction in Crowded Scenes
- Vid2Param: Modelling of Dynamics Parameters From Video
- Composing Diverse Policies for Temporally Extended Tasks
- Physics-as-Inverse-Graphics: Unsupervised Physical Parameter Estimation from Video
- Disentangled Relational Representations for Explaining and Learning from Demonstration
- Hybrid system identification using switching density networks