Microsoft is tryingaccording to TechRepublic, that computers understand what is involved in both the causes of something happening and the effects of actions. Artificial Intelligence systems are not yet very good at it, but if a substantial improvement is achieved, it can result in the tools that use AI being able to help humans make better decisions.
To do this, systems must understand why there are various factors that help people make one decision or another. For example, if it’s hot and a person wants to cool down and is hungry, they are likely to buy an ice cream. Or that if a person has a specific disease but certain characteristics, both physical and psychological, it may be better to apply a specific treatment, and more appropriate for their personal circumstances than the general and more usual one.
For this reason, what the Redmond people want to do is teach an Artificial Intelligence system to decide which is the best question to ask next based on what happened or the answers to previous questions. This will improve decision making Cheng Zhang, Microsoft Research Director.
Zhang notes: “Let’s say you want to rate something, or get information on how to properly diagnose something or classify an item: the way to do it is what I call Best Next Question. But if you want to do something, you want to do it better. You want to give students new material so they can learn better, and you want to give a patient treatment so they can get better. I call it the Best Next Action. And for all this, both scalability and customization are important.«.
To advance this aspect, Zhang and his team have developed a decision optimization API, which combines different types of machine learning to handle both types of decisions in what he calls end-to-end causal inference. This researcher points out that she believes they are «the world’s first team to cover causal discovery, causal inference, and deep learning at the same time. We allow a user who has data to figure out the relationship between all these variables, like which one calls which. And then we understand their relationship. For example, how much of the medicine you gave improved someone’s health, how much a topic will increase a student’s overall understanding«.
The researcher has pointed out that in her research they use «deep learning to answer causal questions, suggest what may be the best next action, in a truly scalable way, that can be used in the real world«.
Companies routinely use AB testing to guide important decisions, but Zhang points out that it has limitations, because “you can only do it at a high level, not at an individual level. You can know that for a given population, in general, treatment A is better than treatment B, but you cannot say which is better for each individual. Sometimes it’s extremely expensive and time consuming, and in some cases you can’t do it at all. What we are trying to do is replace the AB tests«.
Cause effect with Best Next Question
The API that allows to do itcall Best Next Question, is available for private testing on the Azure Marketplace, so organizations that want to use the service need to contact Microsoft first. For data scientists and machine learning experts, the service will be available either through the Azure Marketplace or as an option in Azure Machine Learning. Possibly also as one of the packaged cognitive services, prepared just as Microsoft offers services such as translation and image recognition. Its name could also change to adopt a more descriptive denomination, and indicate that it is capable of performing decision-making optimization.
Microsoft is already exploring its internal use for sales and marketing, starting with the various partner programs it offers. The researchers are also looking at how it could be used in Microsoft Viva, which they are working on with his team. And according to Zhang, they have “many engagement programs to help Microsoft partners grow.” But what they really want is “to discover what kind of interaction program is the treatment that helps a partner to grow more. That’s why it’s a causal question, and we need to do it in a custom way«.
«We want the training to be a personalized scenario: we want people to be taught with the material that helps them the most in their work. We want people to have an intuitive way to use it. We don’t want people to have to be data science experts to do this. Before the end of the year we will release a demo version for end-to-end causal inference«.
This researcher points out that, in the long term, professional and business users will benefit from the systems they already use, such as Microsoft Dynamic and the Power platform, thanks to this. «People who make general decisions need it to be something very visual: a no-code interface where I load the data, click a button and see what the information is.«.
Once causal AI is achieved, a system with correction in both directions can be developed, where humans teach the AI what they know about cause and effect, and the AI can check if it is true. Zhang points out that “if we use the data, we discover the causal relationship and show it to humans«.
But getting end-to-end causal reasoning is only a first step. There is still a lot of work to be done to make it as reliable and accurate as possible. And Zhang is delighted with the potential that she has: «40% of the jobs in our society involve decision-making, and we need to make high-quality decisions. Our goal is to use Artificial Intelligence to help in decision making«.
