Artificial General Intelligence
Of the eight teams competing to develop the first
artificial general intelligence, only one advanced.
The likely reason is that teams must show a plausible
means of successfully completing their grand challenge, and establishing a plausible pathway to AGI
within the timeframe of the competition is itself a
grand challenge. The one team advancing from this
category trimmed their ambitions to a sufficient
degree so that they can plausibly produce their system within the competition timeframe.
Brain Modeling and Neural Networks
Finally, many teams proposed to develop new
approaches to neural networks. These teams often
emphasized architectures that are inspired by the
human brain. While some of the approaches may
prove successful in the fullness of time, there is no
shortage of proposals for new neural network architectures. Without a demonstrated capacity for solving
a problem that was not solvable by previous neural
network architectures, new proposed architectures
36 AI MAGAZINE
Online education is growing rapidly, despite low student
retention for many online classes. The quality of online learning is questionable in part because of a lack of learning assistance. How can we provide meaningful learning assistance to
tens of millions of students taking online classes? Team
emPrize is developing a suite of virtual tutors for online education that mimic many of the roles of human teachers.
These virtual tutors include more than 100 cognitive tutors
for a Georgia Tech online class on artificial intelligence as well
as a virtual tutor for automatically answering questions on
the discussion forum for the class. Preliminary results indicate that student self-efficacy in the class is high and that
interaction with the virtual tutors leads to enhanced student
engagement. emPrize is now expanding the scope of their
work from online education to blended learning; from cognitive tutoring and question answering to exploration and
experimentation, literature survey, and question asking; and
from a class on artificial intelligence to Georgia Tech classes
on introductory computing and introductory biology.
The team contact is Ashok Goel ( email@example.com).
Team Erudite AI
Students who regularly receive private tutoring
score two standard deviations higher on standardized tests than those students without private tutoring. However, the demand for private
tutoring far outstrips the supply, with up to 65
percent of students seeking sessions in Kenya
and 73 percent in Sri Lanka. Consequently,
tutoring suffers from low access, compromised
quality, and the high cost for one-on-one sessions. Erudite AI’s solution endeavors to mitigate all three problems with a peer-to-peer
tutoring platform, ERI (educational real-time
interface). ERI is a human-in-the-loop dia-logue-based tutoring platform comprising
three main components: a mapper to identify
and build a knowledge map of the students’
skills, a matcher to match students to peer
tutors according to their needs, and an amplifier that elevates the quality of the tutoring by
suggesting AI-generated responses for the peer
tutor. In the past few months, Erudite AI evaluated the effectiveness of a dialogue recommender to positive results. Following the
experimental evaluation, the team is producing a scalable open source solution to maximize impact.
For more information, see eri.ai. The team
contact is Hannah Cowen ( firstname.lastname@example.org).
Globally, crop disease causes nearly 50 percent of the total
loss of crops. It is especially devastating for communities in
developing nations where 75 percent of the population relies
on agriculture for their livelihood. Early detection is critical
to fight plant pathogens, as there is a narrow timeframe in
which to intervene to save crops and prevent epidemics.
However, effective early warning systems to alert communities of imminent threats of disease do not currently exist in
DataKind, a nonprofit that uses AI to address complex
humanitarian issues, is developing a model using high-reso-lution satellite imagery at 5 meters per pixel, combined with
computer vision and remote sensing techniques, to detect the
spatial and spectral signature of wheat crops and wheat disease, to be able to provide real-time information on crop disease and support the creation of enhanced early warning systems.
DataKind first worked to identify wheat in Ethiopia, beginning by locating croplands in the region with high spatial resolution. They then successfully built a U-Net model with a 5-
meter resolution to detect croplands in Montana, a climate
proxy for Ethiopia, achieving approximately 93 percent test
accuracy, and a characteristic curve approaching 96 percent
for the area under the receiver operator. The model was transferred using field survey data from Ethiopia, and from human
inspection, appears quite promising. In the second phase of
the project, DataKind is looking to obtain noncrop survey
ground truth data for Ethiopia to further tune and test the
For more information, see datakind.org.