IEEE Circuits and Systems Magazine - Q3 2022 - 52

CAS in the World
4th IBM IEEE CAS/EDS AI Compute Symposium (AICS'21)
T
he 4th IBM IEEE CAS/EDS AI Compute Symposium,
known as (AICS'21), was held over two days
(Oct 13-Oct 14, 2021). The event was very well attended
and received great responses from the audience
all over the world. The symposium was also an initiative
supported by IBM Academy of Technology (https://
www.ibm.com/blogs/academy-of-technology/). Dr. Joshi
has been the main interface for CAS and EDS for organizing
this successful event. This is the second time the
event was organized as a virtual symposium. The audio/
video presentation on this virtual platform went smoothly.
More than 2400 viewers over two days, participation
from 50 countries, over 54 student posters, best
paper poster awards, excellent panel discussions, 11
distinguished speakers from industry and academia
were the salient features of this symposium. There
were more than 5200 views on the LinkedIn post about
the symposium. The theme of the symposium was " From
Ground up to Cloud " . In short, the symposium covered a
range of topics from device technology, to circuits, architecture,
algorithms and sustainability-to make innovations
for the cloud with an emphasis on green AI.
Prof. Hoi-Jun Yoo (Professor of School of Electrical
Engineering and the director of the System Design at
KAIST, Korea) opened the symposium with his excellent
presentation related to " Training on Chip-Next Wave of
Mobile AI Accelerators " . Most mobile Deep Neural Network
(DNN) accelerators target only inference of DNN
models on edge devices, whereas on-device training
was out of reach in mobile platforms due to its excessive
computational requirements. Training-on-Chip (ToC)
with user-specific data is becoming more important
than ever because of privacy issues and communication
latency of training on remote servers. He highlighted
a number of approaches in realizing ToC. General purpose
hardware and software co-optimization techniques
aiming to maximize throughput and energy-efficiency
of DNN training were brought out with examples, such
as, sparsity exploitation and bit-precision optimization
for training. In addition, application specific training accelerators
for Deep Reinforcement Learning (DRL) and
Digital Object Identifier 10.1109/MCAS.2022.3189894
Date of current version: 5 September 2022
52
IEEE CIRCUITS AND SYSTEMS MAGAZINE
Generative Adversarial Network (GAN) were discussed,
touching on issues regarding system implementation
with the fabricated silicon.
Dr. Teo Laino (IBM Distinguished Research Staff
Member) followed up with a very interesting talk about
" A Cloud-based AI-driven Autonomous Lab " . One of the
most significant outcomes of chemistry is the design
and production of new molecules. The application of
domain knowledge accumulated over decades of laboratory
experience has been critical in the synthesis of
many new molecular structures. Nonetheless, most synthetic
success stories are accompanied by long hours of
repetitive synthesis. Automation systems were created
less than 20 years ago to assist chemists with repetitive
laboratory tasks. While this has proven to be very effective
in a few areas, such as, high-throughput chemistry,
the use of automation for general-purpose tasks
remains a tremendous challenge even today. Automation
necessitates that chemistry operators write different
software for different tasks, each of which codifies
a specific and distinct type of chemistry. Meanwhile,
in organic chemistry, Artificial Intelligence (AI) has
emerged as a valuable complement to human knowledge
for tasks such as predicting chemical reactions,
retrosynthetic routes, and digitizing chemical literature.
Dr. Laino's talk highlighted the first cloud-based
AI-driven autonomous laboratory implementation. The
AI assists remote chemists with a variety of tasks, including
designing retrosynthetic trees and recommending
the correct sequence of operational actions (reaction
conditions and procedures) or ingesting synthetic procedures
from literature and converting them into an executable
program. The AI self-programs the automation
layer and makes decisions on synthesis execution using
feedback loops from analytical chemistry instruments,
with supervision from synthetic chemists. He presented
the AI core technology and how it performs across different
types of synthetic tasks.
Subsequently, Mr. Gunnar Hellekson (Vice President
at Red Hat) gave an exciting talk about an open approach
to AI and its integration into cloud. Business innovation
is driven by big ideas, moving faster than ever
before. Today, we can do things we could only dream of
a few years ago. Massive global changes are shifting the
THIRD QUARTER 2022
http://www.ibm.com/blogs/academy-of-technology/ http://www.ibm.com/blogs/academy-of-technology/
IEEE Circuits and Systems Magazine - Q3 2022

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q3 2022
