FIRST ORDER
Learning answers from data
It learns the task it is given. When the task changes, more training may be needed.
SECOND ORDER · Teachable intelligence
Intelligence
you can teach.
People explain. Intelligence learns.
We study how intelligence can abstract experience into concepts and learn from human explanation, not only from data.
Our direction
We build intelligence that learns how to learn.
Sharing concepts with people, one at a time.
We study how intelligence can learn concepts from people’s explanations and examples and use those concepts to solve new problems.
The problem
Experts in the field should be able to teach intelligence.
Today, getting an intelligent system to handle a new task requires substantial engineering work to prepare data and develop the system.
People learn a new concept from an explanation and a few examples. They use it on other problems, and revise their understanding when conditions change.
Experts understand the principles behind their work and the exceptions to its rules. We believe they should be able to teach intelligence through explanations and examples.
Data-centered training Learning from explanations and examples Research goal
How it learns Large-scale data collection and retraining An expert’s explanation and a few examples
New concepts Needs additional training data Learned from explanations and examples
When rules change Requires a review of the data and training process Adapts with limited experience
Who teaches Engineers who work with data and models Experts who know the work
Knowledge representation Inside the model and hard to inspect The supporting evidence can be reviewed and corrected
What we believe
What we believe
- 01
It should be able to learn from limited experience
We examine whether intelligence can learn concepts from a few explanations and examples and apply them to unfamiliar problems.
- 02
People should be able to teach it through explanations and examples
Experts use explanations and examples to describe how their work is done and when exceptions apply. We aim to reduce the effort needed to turn that knowledge into training data.
- 03
People should be able to review and correct what it has learned
People should be able to see what it has learned and which explanations and examples it learned from. They should also be able to correct misunderstandings by explaining again.
- 04
We assess the value of research in practice
Our standard for research is to evaluate learning ability through public evaluations and assess usefulness in real work.
The name
Second order: learning how to learn.
SECOND ORDER
Learning how to learn
The goal is to improve how intelligence learns so it can learn new tasks with limited experience.
SECOND ORDER focuses on research into improving how intelligence learns.
Research
Human-like Conceptual Learning for General Intelligence
We combine fuzzy logic, ontologies, and language models toward intelligence that learns from little experience and stays robust when the rules change.
01 · LEARN
Concept learning
We study how intelligence can learn new concepts by interacting with people. The goal is conceptual understanding beyond memorizing answers.
02 · TRANSFER
Applying what it has learned
We study how intelligence can apply what it has learned to new problems. The goal is to extend that learning to other tasks.
03 · ADAPT
Adapting to change
We study how intelligence can adapt when rules change, using limited experience. The goal is to reduce the burden of retraining.
The path
Learning from real work and expanding into unfamiliar domains
Now
Concept learning research
We study how intelligence can learn concepts from a few examples and apply them to new problems.
Next
Intelligence for work that learns from experts
Building on our experience with projects in the field, we intend to develop tools that experts can teach directly.
Goal
General intelligence
We aim for intelligence that learns new concepts from people even in unfamiliar domains, and connects them to solve broader problems.
In the field
Intelligence for work that experts can teach directly
Practical knowledge is spread across experts’ experience, reports, and test results. We aim to organize the criteria experts use to make decisions and the procedures they follow into knowledge backed by evidence and reusable workflows.
Knowledge
We aim to organize the information in documents into claims, relationships, and supporting evidence.
Work procedures
We aim to make work procedures easy to review and revise.
Adding and revising knowledge
We are developing ways for experts to add and revise knowledge through conversation.
Reviewing supporting evidence
We are developing ways to show the source evidence for each answer and identify what remains unverified.
Founder
Jaehwan Seol 설재환
Developed autonomous driving technology (ADAS) at Hyundai Motor’s R&D center, then led development at several deep-tech startups. Now researches intelligence that abstracts experience into concepts and learns from human explanation, and works on enterprise AI transformation consulting and engineering.
Contact
If you’d like to discuss working together,
leave us a message.
Alongside our research, we provide business consulting and work on AI transformation (AX) projects.