SAN DIEGO, June 20, 2019 /PRNewswire/ -- Lately, the tenth ASU+GSV Global Educational Technology Summit was held in San Diego, CA, USA. Co-founded in 2010 by the Arizona State University (ASU) and Global Silicon Valley (GSV), the ASU+GSV Summit aims to improve learning and career outcomes by promoting innovations on a large scale. It is a high-level grand event on educational technology in the USA and even the world.
Derek Haoyang Li, the founder of Squirrel AI Learning by Yixue Group, was invited to attend this year's ASU+GSV Summit, where he made comprehensive discussions on educational technology innovations with more than 4,500 educational technology experts from all over the world. The Squirrel AI Learning team organized two workshops during the summit for an in-depth discussion on the future development and innovation of AI+ education with other participants at the ASU+GSV Summit.
AI Education Standardization
The first workshop was held to discuss the opportunities for standardization of AI in education based on the IEEE's adaptive instructional system.
The attendees of the workshop include:
At the workshop, the four speakers presented and discussed the current status of the IEEE-related adaptive instruction system, including its design, standardization and best practices. They analyzed the opportunities for applying standardization components and procedures, including AI adaptive learning technology, AI-based recommendation engine and a machine learning-based system used to simulate students' interactions and preferences to improve their learning outcomes, to the AI education system.
Richard Tong: AI redefines the Education Industry
As an addresser, Richard Tong, chief architect of Squirrel AI Learning, gave a speech on the significance and implications of applying AI to education.
In Richard's view, educational technology influences the industry in the following four technical aspects:
Richard further explained, "In our opinion, what AI is to do is redefine. Redefinition means that we are actually changing the educational paradigm, and the whole process has in effect been changed. Therefore, we don't aim to develop a learning model like an assembly line, but to make an individualized model that is truly extendable. We have seen in our work the significance of AI to education, and that's why we have come with all commitments for standardization. "
According to Richard, the successful implementation of AI calls for four elements: business model, professional knowledge, data and computing. Computing power is a broad concept, reflected not only in the number of GPUs and cloud services, but also in stacks and basic computing facilities.
Robby Robson: Why must the adaptive instructional system (AIS) be standardized?
Robby Robson, Member of Board of Governors of IEEE Standards Association (IEEE-SA) and former Chair of IEEE AIS Workgroup, gave a speech on the significance and implications of standardizing the AIS.
According to Robby, there are three important goals of standardization.
Robby also gave a brief presentation of the background of IEEE. As a huge global organization, IEEE has more than 450,000 members and many associations for different discipline categories. Robson's employer IEEE-SA is a global agency made up of more than 7,000 members and more than 300 corporate members. IEEE-SA focuses on industrially recognized standards, and it is concerned about their application to the market and market interdependence, and provides relevant legal and ethical protection.
Avron Barr: The key to the implementation of AI lies in data reorganization and integration
Avron Barr, Chair of IEEE Learning Technology Standards Committee, gave a speech on the data of the AIS. He said, "The market must adjust itself to make the AIS and other AI AI-enhanced products economically viable, but the key issue is: where is the data required for all these AI systems?"
To support an open, AI-based ecosystem, the whole society must build new basic facilities to recollect, share and integrate data. The data comes mainly from five main areas: identity management (personal information), qualifications (capabilities), history (what was done in the past, and what needs to be done subjectively), activity data (e.g., the information on social media, geographical information, etc.), and content. All relevant data owners, such as schools, governments, school districts and businesses, need to get involved to share their data with others to form a large-scale database.
According to Avron, the IEEE Learning Technology Standards Committee has got involved in the establishment of many standards, including the SCORM (Sharable Content Object Reference Model) ten-year update, xAPI 2.0, student data governance and federal machine learning.
Bob Sottilare: Adaptive System Standards and Practices
The adaptive system is a computer-based program that can be used to guide learning in accordance with every learner or team's objective, demand and preference. It recommends an intelligent tutoring system, intelligent instructor and intelligent medium.
As the Chair of IEEE AIS Workgroup, Bob welcomed educators to his work to develop conceptual approaches, set interoperability criteria and recommend practical methods to teachers.
At the workshop, Richard Tong introduced some of the activities and conferences on AI and education to be held this year:
ICICLE 2019 (2019 ICICLE Conference on Learning Engineering) will be held in Ellington, Virginia from May 20th to May 23rd. The conference will focus on learning engineering, make an in-depth discussion on learning, knowledge-based learning experience design, learning conditions and support technologies, and touch upon how the emerging learning engineering can solve the issues of privacy and personalization.
AIAED 2019 (2019 Artificial Intelligence + Adaptive Education) will be held in Beijing from May 24th to 25th. This international forum, held for academic researchers and industry leaders, will focus on AI and other important technologies applied to the next generation of education and how these technologies will influence human adaptive learning.
AIED 2019 (International Conference on Artificial Intelligence in Education), will be held in Chicago from June 25th to 29th. Themed by "Education for All People in the 21st Century", and the conference will focus on how AI and advanced technologies can be used to implement equal education among all people.
IJCAI 2019 Workshop: AI-based multi-modal analysis, used to understand human learning at realistic educational environment workshops.
Knowledge Structure-based AIS
The speaker at the second workshop was Dr. Dan Bindman, Chief Data Scientist of Squirrel AI Learning.
Dr. Bindman graduated from Mathematical Behavioral Science Institute, University of California-Irvine with a doctor's degree in 2002. Later, he joined ALEKS, a leading provider of online adaptive learning in the USA, concentrating on mathematics and chemistry, and became the editorial director and chief architect of ALEKS math products. He joined Squirrel AI Learning in 2018.
Dr. Dan Bindman's speech was titled "Understanding the Advantages and Disadvantages of the Knowledge Structure-based Adaptive Instructional System, and New Ways to Address the Disadvantages" and divided into two parts:
First, he pointed out the advantages and disadvantages of the knowledge structure-based adaptive learning model, and expounded the knowledge point structure and knowledge point state, as well as their differences and the logical links in the knowledge point structure.
Then, he highlighted the PKS optimization model he had developed. Composed of multi-layered student ability groups and more than 20,000 model parameters, the PKS model can accurately judge individual students and groups' real-time learning status and provide personalized learning content for each of them. It can remove the weaknesses of many similar systems with all their advantages retained.
What is a knowledge structure?
A knowledge structure is essentially a bridge among all problems in a course. For example, a typical algebra course may contain 500-1,000 problems (i.e., "knowledge points"). One problem may involve an "equation", while another may involve the "calculation of a slope".
Problems that are highly related to each other can be linked together. If all such problems are linked together in one course, they can form a knowledge structure, which plays a key role in effectively assessing students' knowledge and learning capacity.
What is a knowledge state?
Students' knowledge states can be expressed as a vector that represents the knowledge or their mastery of all problems in a course. It is worth noting that Students' knowledge states are binary for all problems, i.e., "known" or "unknown", or seen as "1" or "0".
For example, a course contains 1000 problems, and students' mastery of knowledge can be expressed as {0,1,1,……,1,1} through the vector.
What's the relationship between the knowledge states and knowledge structure?
It depends on whether there is a connection among different problems. Dr. Bindman enumerated two extreme possibilities: "completely unconnected" and "completely interconnected".
Suppose there are 3 problems in a course. If the problems are completely unconnected, the likelihood that a student masters these problems is equal to "3 squared", i.e., 8, ({0,0,0}, {0,0,1}…{1,1,1}).
However, if the three problems are interconnected with difficulty increasing progressively, it means that the student will have little difficulty solving easy problems if he overcomes the difficult problem. The number of his knowledge points can be reduced to 4: ({0,0,0}, {1,0,0}, {1,1,0}, {1,1,1}).
By analogy, we can guess that there will be fewer knowledge states if problems are more interconnected; on the contrary, there will be more knowledge states if problems are less interconnected.
So Dr. Bindman asked another question: how many questions do you need to ask to understand a student's knowledge states?
Based on the above conclusion, if problems are completely unconnected, you need to ask 1,000 questions to get the answer; on the contrary, if problems are related to each other with difficulty increasing progressively, you just need to ask 10 questions: Ask the 500th question first. If the student gives a correct answer, continue to ask him a more difficult ones: the 750th; if he fails to give a correct answer, continue to ask him an easier one selected from the medians.
How to apply the knowledge structure to adaptive learning?
In view of Dr. Bindman's introduction, the knowledge structure links different problems together, and a student must master some essential knowledge before beginning to learn a particular problem. Besides, the knowledge structure can greatly reduce the number of questions asked to assess the student's mastery of knowledge. Of course, an issue will therefore arise: the knowledge structure must contain a lot of connections if it is used to assess a student's knowledge states.
The connections can also be divided into different types: logical connections and experiential connections. The former is based on the logical relationship among problem contents; the latter is based on the correlation of probabilistic data among problems. The above is a conclusion drawn based on data collected. For example, a student who can solve problem A can also solve problem B. That's an experiential connection.
Why is an experiential connection important? This is because logical connection features sparse data, and you don't have enough experts to do manual annotations. An experiential connection can generate a lot of data, although problems still arise, because some experiential connections are inconsistent in sequence with textbook knowledge.
Another challenge posed by the knowledge structure comes from forms of curriculum. If a student guesses right a lot of questions, there will be something wrong with the knowledge structure-based assessment. Relatively speaking, the knowledge structure is applicable to courses in mathematics, chemistry and physics because there are many fill-in questions in such courses; on the contrary, it does not apply to courses in Chinese or English due to a large number of choice questions.
The last challenge comes from nature of the knowledge structure itself. Mastering knowledge is not absolutely a binary problem. For some knowledge, you can solve problems even if you just master 80% or 60% of it. In this case, the knowledge structure cannot accurately assess whether you have mastered the relevant knowledge points.
In summary, Dr. Bindman drew the following conclusion: in the adaptive learning area, the knowledge structure can be a very powerful tool, but there remain lots of difficulties to overcome:
To solve these problems, Dr. Bindman has designed a model: PKS.
What is PKS?
PKS is an abbreviation for Probabilistic Knowledge State. Compared with the binary system, expressed as "1" or "0", Dr. Bindman suggested expressing a student's knowledge states with the probability of answering every question correctly at a specific time, such as {0.29, 0.87, 0.63, 0.74……}.
Any student's PKS at any time is entirely determined by his ability in the "knowledge channel". The knowledge channel is a pure mathematical structure during model fitting. Dr. Bindman said that he didn't know what they represented, but after fitting, researchers could be inspired after getting to know that for a problem, which channels were very important, while which were not.
Dr. Bindman said, "You know that the knowledge channel is a block box, i.e., you don't know how it works internally. In a sense, it's similar to principal component analysis."
PKS is determined by three variables: students' ability A(t) to answer questions within the prescribed time t; question weight W(q), i.e., the correlation of the question to the knowledge channel; question core C(q), which can be understood as Bias. The value of PKS is PKS(q) = Phi[A(t)xW(q)+C(q)]], where "Phi" is a normal cumulative distribution function; "." represents vector dot product.
This method can enable you to apply this formula to all problems at any time. You can make parallel use of this equation within time t in order to gain complete PKS for all problems at any time.
Another advantage of PKS is that students' learning history can be fully used. When a student starts a course, he has to develop all his abilities from scratch; as he keeps learning, the model will adjust its parameters by reference to his learning history and then make a more accurate assessment of his abilities.
According to Dr. Bindman, Squirrel AI Learning by Yixue Group has started applying the PKS model to middle school mathematics teaching.
SOURCE Squirrel AI Learning
