Streamlining Learning Curves in the Workplace with Machine Learning

by Avery Anderson

Photo credit: HbrH/Getty Images

Working Smarter, Not Harder

Starting at a new job can be tough. Learning the ropes involves more than just mastering information and understanding the tasks you are expected to complete. Many times, there are tips and tricks that are role or company-specific. 

These helpful hacks are harder to learn because they are not always explicit. Instead, they are often found in the unique experiences and knowledge that only someone who worked the job before has, which can get lost in the transition from old to new employee. 

After a few weeks or months at a new job, you gather enough information to be successful, but you might have to create your own methods for completing tasks or even consult the person who had the job before you. 

What if you could learn everything you needed to know quicker, without struggling through unhelpful tutorials or having to bother a prior employee or supervisor?

USC Viterbi Information Sciences Institute’s (ISI) Knowledge Needed in Context (KNIC) program is developing a way to get this valuable information out of the prior employee’s head and into a form for the new employee to use.

The Secret Ingredient–AI Helpers

The KNIC program is a part of the Knowledge Management at Speed and Scale (KMASS) project that the U.S. government’s Defense Advanced Research Projects Agency (DARPA) created. (1)

Jay Pujara, Director of the Center on Knowledge Graphs at ISI, said the idea behind KMASS was building AI that can “interface and watch somebody do their work, learn from it in the way an apprentice might, and then store this knowledge so that it can then be provided to the new person and help them learn the job.”

Thinking back through human history, implementing this “old apprentice method” makes sense–shadowing someone who has mastered the job makes getting over the learning curve easier, because you gain their own tried and tested methods. 

In this case, the skilled employer is in the form of an AI “companion” who is able to evaluate what the novice is doing and what they need help with, without annoying them or stunting their productivity. 

“The core research vision we have is building this companion that has dialogue with you, either a knowledge producer, doing a task and it’s watching you and trying to learn from what you did, or as a knowledge consumer and it’s watching you and it’s trying to give you some helpful suggestions,” Pujara explained.  

Take typing a line of code, for example. This AI companion would be able to identify in real time when the person is making a mistake, and suggest a fix or provide guidance on where they could go to find help. 

It’s All About Knowledge

KNIC is broken up into three distinct pieces: knowledge store and organization, knowledge capture, and knowledge dissemination. 

Knowledge store describes the process of compiling background knowledge about a topic, pulling from lab and academic documentation, the internet, and other sources to establish context. 

Knowledge capture refers to the AI companion’s interactions with a knowledge producer, learning what they are doing and figuring out what questions to ask and when to ask them during the task demonstration without redundancy and irritating the person using the system. 

The third area, knowledge dissemination, is communicating and delivering information to the knowledge consumer that would be useful in the context of what they are trying to accomplish and based on what they already know. 

Elizabeth Boschee, Associate Director of the AI Division at ISI and Director of ISI Boston, said the team began their research by experimenting internally.

“We first had a masters student at USC go through the documentation and try to build a model,” Boschee explained. “She was unsuccessful, which was in this case good because it shows that the documentation produced by our own labs is imperfect or it ages poorly.”

This “guinea pig” experiment helped the team learn what kind of information would have been useful for the system to provide the student when she got stuck on a step.

But would a new employee really learn how to complete the tasks on their own if the bot was there to give them answers every step of the way? What about the age old classic “learning by doing” method?

Boschee says the KNIC program would actually make it easier to learn by doing because the  technology would not be completing tasks for the user, but helping them find the knowledge they need quickly when they hit a roadblock. In this way, the learning by doing process would be made more efficient. Users are not being fed answers, but help is there if they need it.

Invasion of Privacy? 

This technology sounds like a handy tool for a new employee, but what does the documentation process look like for a current employee? In this master–apprentice model, is the bot always watching?

The KNIC system would not be constantly monitoring the employee–it would be a choice whether the technology is used or not used in a particular situation. 

“Right now, students who graduate or workers transitioning to a new job have to sit and write down everything they know, and that’s a really hard process. Instead, they could use our system to demonstrate the different tasks that they do and have the system ask questions about the parts that are unclear or represent something new or unknown,” Pujara explained.

In this way, it would be an “assistive tool the same way someone might use Zoom to record a lecture or demo,” Pujara added.

Just the Beginning

KNIC is still in the early stages, in fact, it is only month six of what is intended to be a three year long project. Currently, the team is working on building a functioning prototype for the data science domain. 

One of the future challenges for this project is being able to make it more generalizable. 

“It’s one thing to be tracking someone doing a task on the computer where everything they’re doing is on the computer. Whereas if a cook in training accidentally puts egg yolks in instead of egg whites, that’s a very different kind of error and detection of that is very different,” Boschee said. 

On top of that, machines are still nowhere near as powerful as humans. We still have a long way to go before AI can completely replace the role of a coworker training someone new.

Jonathan May, Research Associate Professor of Computer Science, said even though Al is “very weak relative to what humans can do,” ISI is able to work on “simpler problems that would still increase efficiency and allow for better transfer of knowledge.”

One of the areas that this project is applicable, Boschee said, is in the armed forces. 

“The government cares because this is a very common scenario in a military context because people are on rotation, and making those role transitions without needing access to the previous guy would be incredibly helpful,” Boschee explained. 

Pujara and Boschee are conducting the dissemination system and interfaces. Muhao Chen, Research Assistant Professor of Computer Science, is developing the background knowledge and extraction piece for the knowledge capture.

May, along with David Traum, Director for Natural Language Research at the USC Viterbi Institute for Creative Technologies (ICT), are both working on interface design and building the AI agents that would interact with the human user.

The KNIC project shows a promising future for the efficient transfer of knowledge in the workplace. With an AI companion on your shoulder, tackling a new role may be less scary and more seamless. 

(1)KMASS seeks to create technology that aids in knowledge documentation, acquisition, and application. DARPA chose proposals from various research institutions, one of which being ISI’s KNIC project, which focuses on the data science domain. This fall, the project received a grant for almost seven million dollars

Published on January 31st, 2023

Last updated on January 31st, 2023

Want to write about this story?