当前位置：网站首页>Design artificial intelligence products: technical possibility, user acceptability and commercial feasibility

Design artificial intelligence products: technical possibility, user acceptability and commercial feasibility

2022-06-28 13:28:00 【Liguodong】

With machine learning becoming the mainstream of digital products , Understanding the basics of machine learning is becoming increasingly important for many product managers . Today's product personnel are a fairly diverse group . For some people , The focus is mainly on the user experience （ for example , If the main value proposition revolves around a killer UI）, Others are already designing products that require a deep understanding of data and code .

Understanding machine learning is necessary at both ends of the spectrum —— Just for a slightly different reason . about With UI Centered products and PM, Fuzzy logic and machine learning will fundamentally change the way users interact with products . therefore , The presentation of these features becomes very important . On the other hand , management API Or the product manager of the technology platform Will care more about AI How algorithms are integrated .

Product management, like machine learning, is a huge topic , So let's start with a basic question . When is it worth developing AI products ？ Of course , You can also apply this problem to functionality in the context of a larger product .

majority PM A useful tool to see is IDEO Popularized Sweet Spot for Innovation（ The best point of innovation ）. It discusses the desirability of the product 、 Possibility and feasibility , Valuable ideas tend to touch all these areas . If you are not familiar with this framework , You should Read this article .

Technical possibilities （feasibility）： Can we really do it at this stage , Is it possible? ?
User satisfaction （desirability）： Can it solve customers' problems ? Is it what customers expect ?
Commercial feasibility （viability）： Should we do this ? Will it succeed in the future ?

Let's look at these concepts from the perspective of machine learning products .

Technical possibilities （feasibility）

Possibility is not usually the part I suggest you start with when evaluating product ideas . For all that , Compared with traditional software products , This is probably the most different aspect , Especially because we are still in ML The early stages of enhancing software functionality .

Although time estimation is still difficult , We are all very good at assessing software possibilities . When you describe a problem to a developer , They are already thinking about technology and libraries . You may be considering that previous products have solved the relevant problems . as time goes on , There will be similar thinking patterns in machine learning .

In terms of possibility , The question you should ask yourself and your team is ：

What is the problem we are going to solve ？

For any question , Problem setting is necessary , But this is especially true when dealing with greater uncertainty . for example , Suppose our idea is to detect defective products on the production line . Granularity becomes very important here ： How much improvement do we look for in the defective products we send to our customers ？

Do we have any data on this problem ？ without , Can we get data on this issue ？

Machine learning is all about data . It doesn't always mean big data , It is Sufficient quality data about the problem . Some data is easy to access . for example , User actions from your application . They may already be stored somewhere .

Other data types are more difficult , for example ： A large number of annotation image sets . Check out our product line examples , You need an image set , It includes perfect and defective products , And mark and shoot defects from the same angle that you put the sensor in the production line .

Whether there are patterns in the data that are meaningful to the algorithm ？

The difficulty of machine learning is , Data scientists usually need to do a lot of work first to evaluate data sets and conduct experiments , To see if... Exists in the data ML Models can understand patterns . Different from traditional software , If you don't really work hard , It is difficult to assess the possibility . for example , Use your annotated product dataset （ Defective and flawless ）, The model will be able to distinguish between the two .

In traditional software , The possibilities can almost be described in two parts （ Possible or not ）. However , In machine learning , Technical possibilities may be more of a range , It will overflow to the range expected by the user .

User satisfaction （desirability）

Figuring out what people want is a tricky job , And for AI product , No exception . Assessing acceptability is ostensibly the same as traditional software , But there is a trap . Give Way AI It seems highly desirable to evaluate every product on the product line . in fact , Many companies are studying these types of solutions . But there are other questions you should ask ：

What is the performance of the algorithm and what performance it needs to perform ？

This problem is related to the technical possibility and user satisfaction of the solution . Before data scientists work hard , There is no simple answer .

You may find that the algorithm will detect defective products at a high speed , But it will also produce false positives . What does this mean for your production line ？ Perhaps the production workers will completely ignore the algorithm because of false positives , And your solution becomes undesirable .

It is easy to describe a fully autonomous solution , But it's not always possible to build . Perhaps the product you can actually build is just a human computer assistant . Fully autonomous vehicle and Computer assisted steering In terms of user satisfaction, they are quite different . But it's best to consider that something in your context becomes a desirable point for users , This also relates to the next question .

How much control the user will have ？ Whether users trust your solution and see value in the same way as you ？

Trust is an important topic in artificial intelligence , And for good reason . Machine learning models can be black boxes for those who develop them , So imagine how the end user feels . An important aspect to consider is how much information you must convey to convince users , And how much control you have over users' decisions .

for example , In our production line example , Do you need to show the confidence of the algorithm for each prediction （ That is, the product has 70% Possibility of failure ）？ Will you give the line manager the ability to adjust the confidence that defective products will be discarded ？

This is, of course, a simplification . Giving control of the algorithm to the user may be difficult to implement and difficult for the user to understand .

These issues become extremely important in sensitive scenarios such as healthcare , Because the unintended consequences can be terrible .

Commercial feasibility （viability）

If you come up with a feasible and ideal solution , The question then is how much it is worth to users and ultimately to you . There are many ways to assess commercial viability , and PM Must always consider whether the new product concept is in line with the company's strategy .

Some of the more relevant aspects of feasibility are specific to ML The problem is ：

Will the long-term value outweigh the short-term cost ？

Development ML The cost of functionality can be prohibitively high . Considering that data scientists and machine learning engineers are the most popular talents today . Even if you already have them on your team , They can also take many other steps . Besides , Designers and subject matter experts who are proficient in machine learning may be even rarer . The cost of obtaining high-quality data is also high , And the training model is not completely free .

From data collection to model services , You need to write a lot of code and build the infrastructure . most important of all , Machine learning is new to most people , And all stakeholders are involved in a lot of education and change management .

Due to the cost and expertise required , Maybe it's always worth asking Whether expert system can be used （ Rule engine ） instead of AI To solve some problems . secondly , Ask you How to ensure that your team uses research and technology that others have built before them . These two questions may save you and your company a lot of money .

Will the problem change over time ？ Can the solution be extended to other areas ？

Two other areas will significantly affect your AI The feasibility of the product .

On the one hand, you deploy the solution The dynamic degree of the environment . In our production line scenario , Will the product model change frequently ？ If so , To maintain the viability of the solution , get data 、 Training ML The cost of models and other updates will be overtime . This problem is more common than you think .

The other side may be Your solution can be replicated to other similar problems , You can reuse most projects . for example , You may have originally designed a fault detection solution for a single product . For all that , The same sensor can be installed on other production lines , And you can collect training data . These types of opportunities can justify many significant upfront investments .

Combine them

In machine learning , The product manager （ But not limited to the product manager ） A little naive . Their attitude changed from Completely beyond the scope of possibility ML To Easy to solve all problems ML . So naturally , The truth lies somewhere in between .

As this article shows ,AI User satisfaction in the product 、 Technical possibilities 、 Commercial viability is interrelated , Compared with traditional software , Revealing the relationship between them will require more specific experiments and prototype design . meanwhile , Data scientist 、 The engineer 、 Product managers and subject matter experts need to work together from the start .