Invited speakers

Prof. Kenji Doya, OIST

Title: What can we learn from the brain for cognitive robotics?

Abstract:

Deep learning and reinforcement learning are prime examples of how brain-inspired computation can advance AI and robotics.

Then what else can we further learn from the brain for the future designs of cognitive architecture in robotics?

The aim of this talk is to overview evolutionary, developmental, and meta-learning features of the brain that are relevant for artificial general intelligence and to point to relevant examples and directions in brain-inspired cognitive robotics.

Reference:

Doya K, Taniguchi T (2019). Toward evolutionary and developmental intelligence. Current Opinion in Behavioral Sciences, 29, 91-96. https://doi.org/10.1016/j.cobeha.2019.04.006

Dr. Masahiro Suzuki, The University of Tokyo

Title: Pixyz: a framework for developing complex deep generative models

Abstract:

In recent years, researches on deep generative models (DGMs), probabilistic models parameterized by deep neural networks, have been advanced rapidly in the context of deep learning.

These are also important in the field of robotics because of their ability to model the environment, or our world, directly from multimodal sensor information.

However, since the recent DGMs become more complicated, it is difficult to implement them with conventional deep learning libraries and probabilistic programming languages.

Therefore, we first had a rethink on DGMs focusing on the differences with conventional probabilistic models, and then, based on this rethinking, we developed a library for implementing DGMs, which we call Pixyz.

Our library can more concisely implement complex DGMs, which is difficult with previous libraries.

In this talk, I will explain the background and concept of Pixyz, and give a brief tutorial about it.

Then, I will talk about the future outlook of applying Pixyz to the field of robotics and cognitive science.

Prof. Tomoaki Nakamura, The University of Electro-Communications

Title: A Framework for Construction of Multimodal Learning Models

Abstract:

To realize human-like robot intelligence, a large-scale cognitive architecture is required for robots to understand their environment using multimodal information obtained by various sensors.

So far, we have proposed multimodal learning models that enable robots to learn motions, places, objects, and language.

However, these models have a complex structure and it is not easy to implement them.

As the scale of models becomes larger, it is more difficult to construct such models and to derive and implement the equations needed to estimate their parameters.

To overcome this problem, we have proposed a framework named Serket that enables the construction of large-scale cognitive models by connecting smaller fundamental models hierarchically while maintaining their programmatic independence.

Moreover, the connected modules affecting each other and their parameters are determined by their communication with each other.

Therefore, their parameters are optimized as a whole. In this talk, I will describe the framework Serket.

I will also introduce some implementation examples and show that it is easy to construct integrated models by using Serket.

Reference:

Tomoaki Nakamura, Takayuki Nagai and Tadahiro Taniguchi, “SERKET: An Architecture for Connecting Stochastic Models to Realize a Large-Scale Cognitive Model”, Frontiers in Neurorobotics, vol. 12, article 25, pp. 1-16, 2018 [Link]

"Serket: A library for constructing large scale cognitive model", http://serket.naka-lab.org/

Dr. Igor Mordatch, OpenAI

Canceled

Dr. Douwe Kiela, Facebook AI Research USA

Title: Multimodal machine learning: an NLP perspective

Abstract:

Natural language processing has made tremendous progress in recent years, in part due to self-supervised pre-training and transfer learning.

I argue that while processing large quantities of language data in itself is useful for developing natural language understanding, it is not sufficient - what is missing is information from different modalities that grounds the language.

The talk will go over recent work in grounded natural language processing, with a focus on incorporating multimodal content into representation learning and generative models.

The hope is that recent developments in NLP may be useful for the field of robotics as well.