Blog: How to manage “AI” development and an R&D team in big corporations without all of the “magic” and…
Since my main line of work deals with AI, brains for robots, ( Augmented Pixels, Inc) and I write in a blog about AI use cases in different industries(AI.Business), some big corporations involved in the traditional industries, from time to time, ask me to help them solve problems with their AI research & development.
Usually these problems are connected to an “over-budgeting” of R&D or to problems with meting goals, or to inefficient management in R&D teams.
Here are the most common issues and takeaways from my experience:
#1. “AI” is not magic and should be developed and delivered within time limits and budgets.
Almost all tasks (roughly 90%) that are performed in areas of Computer Vision and Machine Learning are not related to “real” R&D (creating algorithms or significant improvements).
For successful implementation it is usually simply enough to just find a relevant previously published paper and implement it with good quality…
So do not go out and hire “witch doctors”, who will beat their drums and tell you about the “magic” of AI, but can’t tell you when and how the problem will be solved. Hire only the specialists who can in a highly technical way describe the entire process involved in reaching the goal and who do not try to reinvent the wheel instead of simply implementing a well-known correct method.
#2. For the hardest R&D tasks in “AI” people are your main asset, for the other tasks its data.
When implementing the pluses and minuses (+/-) of standard functions, you can hire middle people, whose task will most likely be to create a well-marked and cheap dataset, to implement a well known and described approach and to create a system for correct data collection and subsequent training of algorithms. In this case, the business’s main assets are datasets and the data itself that is created and collected in the process of using the system (i.e. through its additional training, etc.).
If development accounts for 10% of your tasks and its implementation requires real R&D, then your main asset is the people who will implement it. Therefore, the majority of the budget, both of yours and of your competitor’s, will be spent not on implementation, but on testing incorrect hypotheses. And in this situation it is essential knowledge to know precisely which paths lead to results, and which paths do not.
Of course, when creating such projects, you need to reinforce your competitive advantage with an efficient logging system, data collection system, and a subsequent training system.
#3. The hardest part is integrating AI solutions into a business process and starting to use it. The easiest part is developing it.
If a large corporation creates a well-functioning AI solution, then soon they will be shocked to find out that the integration of the solution into their business process COSTS much more than it cost to create the AI system in the first place.
This often occurs due to broken communication between those who are directly involved in the business process and those who innovate.
Encourage AI developers to get out of their lab and into the “field” as quickly as possible in order to interact directly with the future internal users of the AI system.
#4. Most AI solutions are useless, because, internally, people just don’t use them
AI system developers typically do not like to go out into “the real world”. For example, in agriculture often when an AI system is created its implied that the AI developer used “STANDARD machine learning approaches on some pictures and got some kind of result.”
Everything seems to be just fine — an AI system was successfully created after all. But when the end result of a system like this gets into the hands of an expert (for example, an agronomist), it winds up being absolutely useless to him.
Of course, the “let’s go invite an AI witch doctor, who will create an AI system using standard machine learning methods” approach is cheap and can give useful results, but, typically, in my practice, the results are going to be useless. To get really useful results, the AI system developers need to dive deep into the subject area and understand all the nuances in order to create neural networks, which as a RESULT, in the future, will compete with the living neural networks of specialists in this specific field who have studied it as their major for 5–6 years and then practiced it for 10 years.
#5. 50% — 70% of problems/tasks could be solved in a more efficient way without using “AI”
Nowadays, a lot of AI startups are trying to offer silver bullets to do a variety of tasks that have been done very effectively for a long time using well-known methods that do not require AI (simple statistical methods, etc.).
Therefore, before developing any AI system, it is necessary to calculate how much it will cost to create it, as well as the cost of its OPERATION and make certain that it is predictably more effective than currently existing approaches — if the system can only give a forecasted improvement that is within the margin of error, then there is no point in making it.
Instead of a conclusion
In general, I would recommend that large corporations stop “worshiping” AI “witch doctors” and begin applying the same decision-making and management methods to AI projects as they do to others.
There is no “magic” in modern AI (CV / ML, Deep Learning), therefore AI projects should be carried out according to planned budgets and schedules: =)
vactivity @ gmail.com
Originally published at http://vactivity.com on May 14, 2019.