Human-Machine interaction aspects of Machine Learning systems
by Nitin Kumar
A previous blog article introduced the Machine Learning (ML) model lifecycle defining various steps in realizing an ML project. Today our goal is to help you classify the kind of ML use case you are developing in terms of how it will interact with humans. Knowing and following this classification will help you and your company in all steps of a ML model cycle.
Here are the five ways an ML algorithm (also referred to below as “AI”) can interact with people [1].
1. AI decides and implements (“automator” scheme) - In this particular form of interaction, the assumption is that involving humans would only slow down the process; as a result, the ML algorithms do almost all of the work. They must have access to all the data and context. Two examples for which this approach may be appropriate, both applicable to grocery store management, are clearance pricing for unsold and soon-expiring inventory, and personalized customer discounts.
2. AI decides, Human implements (“decider” scheme) - Here AI captures the context and makes the decisions, then Humans will implement the retained solution. In online grocery shopping, if a customer selects out-of-stock products, an ML algorithm can use historical data to suggest alternatives. Humans can check the quality of the suggestions before delivery. As another example, an AI algorithm can identify failures in production facilities, then a human runs the repair. This category can be called a Decider.
3. AI recommends, Human decides (“recommender” scheme) - With Google Maps, an AI algorithm recommends multiple options to reach a destination, then a human chooses one. As another example, an AI algorithm can suggest what products to buy to replenish the stock of a grocery store; if it does not have access to the supply chain, the store manager will make the final ordering decision.
4. AI generates insights, Human does decision making (“illuminator” scheme) - Here insights from AI support the creative side of humans. An AI algorithm for a grocery chain could for example identify shopping patterns unique to geographical locations and use them to produce recommendations for merchants on possible location-specific features. As another example, an AI algorithm can shed light on future workforce needs, on which HR professionals can rely to gain a competitive edge.
5. Human generates, AI evaluates (“evaluator” scheme) - In all previous cases the flow went from AI to humans. Here we reverse it: humans generate hypotheses, AI tests them. The most important example is “digital twins” technology, whereby people at a company produce many scenarios based on a digital model of some asset (such as a manufacturing plant); then AI serves to simulate and assess those scenarios. Other examples of the evaluator scheme involve the testing of rare scenarios: for online stores, assessing the impact of a pandemic such as Covid-19; for disaster relief agencies, assessing the impact of hurricanes using historical data.
Organizations that successfully use ML to drive growth are incorporating all five modes described above in their business processes. For effective user testing and for successful integration of the machine-learning processes into the business operations, it is essential that ML practitioners know in each case which of the five schemes is being applied.
This article will, I hope, help you in this identification process. To acquire an in-depth practical understanding of how to apply data science in industry, join one of our data programs: AI for Managers, Data Science and Machine Learning.
Here are the five ways an ML algorithm (also referred to below as “AI”) can interact with people [1].
1. AI decides and implements (“automator” scheme) - In this particular form of interaction, the assumption is that involving humans would only slow down the process; as a result, the ML algorithms do almost all of the work. They must have access to all the data and context. Two examples for which this approach may be appropriate, both applicable to grocery store management, are clearance pricing for unsold and soon-expiring inventory, and personalized customer discounts.
2. AI decides, Human implements (“decider” scheme) - Here AI captures the context and makes the decisions, then Humans will implement the retained solution. In online grocery shopping, if a customer selects out-of-stock products, an ML algorithm can use historical data to suggest alternatives. Humans can check the quality of the suggestions before delivery. As another example, an AI algorithm can identify failures in production facilities, then a human runs the repair. This category can be called a Decider.
3. AI recommends, Human decides (“recommender” scheme) - With Google Maps, an AI algorithm recommends multiple options to reach a destination, then a human chooses one. As another example, an AI algorithm can suggest what products to buy to replenish the stock of a grocery store; if it does not have access to the supply chain, the store manager will make the final ordering decision.
4. AI generates insights, Human does decision making (“illuminator” scheme) - Here insights from AI support the creative side of humans. An AI algorithm for a grocery chain could for example identify shopping patterns unique to geographical locations and use them to produce recommendations for merchants on possible location-specific features. As another example, an AI algorithm can shed light on future workforce needs, on which HR professionals can rely to gain a competitive edge.
5. Human generates, AI evaluates (“evaluator” scheme) - In all previous cases the flow went from AI to humans. Here we reverse it: humans generate hypotheses, AI tests them. The most important example is “digital twins” technology, whereby people at a company produce many scenarios based on a digital model of some asset (such as a manufacturing plant); then AI serves to simulate and assess those scenarios. Other examples of the evaluator scheme involve the testing of rare scenarios: for online stores, assessing the impact of a pandemic such as Covid-19; for disaster relief agencies, assessing the impact of hurricanes using historical data.
Organizations that successfully use ML to drive growth are incorporating all five modes described above in their business processes. For effective user testing and for successful integration of the machine-learning processes into the business operations, it is essential that ML practitioners know in each case which of the five schemes is being applied.
This article will, I hope, help you in this identification process. To acquire an in-depth practical understanding of how to apply data science in industry, join one of our data programs: AI for Managers, Data Science and Machine Learning.
