Date. 20 Jul 2017 (Thur)
Room. 1.502-36 (Meeting Room 10)
Speaker. Nengli Lim (SUTD)
Title. Rough Paths and its Applications in Machine Learning
Slides. Click here.
Some papers that might be useful:
Here are some information of some conferences:
If you have any conference that you would like to suggest, please leave a reply in the comments section below. Thanks!
The NVIDIA DGX-1deep learning supercomputer is winning over universities in Singapore. Singapore University of Technology and Design (SUTD) and Nanyang Technological University (NTU) have deployed the powerful machine for their research projects on artificial intelligence (AI).
SUTD will use the DGX-1 at the SUTD Brain Lab to further research into machine reasoning and distributed learning. Under a memorandum of understanding signed earlier this month, NVIDIA and SUTD will also set up the NVIDIA-SUTD AI Lab to leverage the power of GPU-accelerated neural networks for researching new theories and algorithms for AI. The agreement also provides for internship opportunities to selected students of the lab.
“Computational power is a game changer for AI research, especially in the areas of big data analytics, robotics, machine reasoning and distributed intelligence. The DGX-1 will enable us to perform significantly more experiments in the same period of time, quickening the discovery of new theories and the design of new applications,” said Professors Shaowei Lin and Georgios Piliouras, Engineering Systems and Design, SUTD.
Commentary by L. RAFAEL REIF (President of MIT)
One of the co-founders of Geometric Intelligence that Uber bought over, is Zoubin Ghahramani, a leading expert on Gaussian processes.
We are still pretty far from scalable common sense in AI.
Date. 11 Nov 2016 (Fri)
Speaker. Shaowei Lin (SUTD)
Title. Deep Probability Flow
Abstract. I will be speaking on our recent work on Deep Probability Flow. This is an algorithm that we hope will enable us to train much larger neural networks. Current methods in deep learning are hindered by back-propagation which computes the maximum likelihood estimate. Such algorithms may be distributed using data-parallelism where the training set is partitioned between several machines for gradient computation. The holy grail, however, is to achieve model-parallelism where different parts of the neural network are trained on separate machines. By using a different statistical objective function, we were able to derive learning rules where the weight updates only depend locally on the associated neurons. In the talk, I will discuss how variational inference is used to derive these rules. This is work in progress.
Other people have already started noticing the relationships to Homotopy Type Theory! =)
Incidentally, here are the beginnings of a categorical approach to agent interactions!