President, Japan Deep Learning Association (JDLA).
Professor, Department of Technology Management for Innovation (TMI), Graduate School of Engineering, and Program for Social Innovation (PSI), School of Engineering
Webpage: http://ymatsuo.com
World Models
World model in human brain —human can not perceive everything.
- The brain models how the world evolves through time, locations and events.
- The brain unconsciously predicts the world according to the model
We are focusing on deep generative models to build a world model. The questions we are asking throughout the research are:
- How can we model a complex world efficiently?
- How can we combine a world model and language? How can language or symbolic information ne integrated into World Models.
1. Fundamentals of AI, ML and Deep Learning for Product Managers
2. The Unfortunate Power of Deep Learning
3. Graph Neural Network for 3D Object Detection in a Point Cloud
4. Know the biggest Notable difference between AI vs. Machine Learning
Joint Multi Model VAE
Robots acquire various types of information from multiple sensor; video, audio, angle, acceleration, etc. Our goal is to make it work to integrate them (different modalities) for better prediction and decision making.
In this work (Joint Multimodal Learning with Deep Generative Models) we build the encoder for each modality and learn them to approximate the original VAE encoder. After finishing the training, we can use each trained encoder (for each modality) to infer properly from a single modality.
We can obtain joint representation and perform bi-directional generation — images and attributes. The idea is applied to semi-supervised learning and zero-shot learning: Semi-supervised Multimodal Learning With Deep Generative Models.
There are various deep learning frameworks for implementing the deep neural networks. Since these frameworks are not treated as probabilistic models, it becomes difficult to implement complex deep generative models.
Fortunately, there are some extensions of these frameworks such as TensorFlow Probability, Edward, Pyro. We have developed a PyTorch based library for deep generative models: Pixyz.
Offline Reinforcement Learning: improving a policy under a pre-given trajectories. The assumption for offline RL is that there is no interaction with environment during the policy improvement. And the merit is to use of RL for problems where it is difficult search for measures in the real environment such as medical care, education, dialogue agent, real robot control.
We proposed a new concept of deployment efficiency: Deployment-Efficient via Model Based Offline Optimization.
This research open a direction which is learning a world models from others’ experiences. Such approach might be necessary to model the complex world.
Credit: BecomingHuman By: Alisher Abdulkhaev
