(This is an article I translated from the article Moving AI from PoC Stage to Production by Alexandre Gonfalonieri. In the article below, I will translate the first part: The actual implementation of the PoC project and the reason why the PoC project failed. Pilot phase and convert PoC into products will be further translated in the following article.)
After working on various Artificial Intelligence (AI) projects, I realized that most PoCs (Proof of Concept) have not reached production stage and only a handful have come. release phase. In 2019, many companies have started to apply AI solutions with impressive results but similarly, only a few businesses have the ability to fully develop AI capabilities and bring added value. Rise for his organization. Based on my own experience, I find that less than 20% of all PoC projects that apply machine learning reach production. And even up to this stage, most projects will fail in the “industrialization” of AI solutions.
1. Actual situation:
Most companies start by demonstrating that an AI solution will, in fact, help cut costs, improve customer experience, and make a difference for businesses, through a Proof of Concept. (PoC). PoC is usually implemented on fairly simple algorithms using existing training data or internal data. The main purpose is to show that an algorithm can be trained to solve a specific case with only a small amount of training data. And if they succeed, the PoC project will continue into the production stage. In fact, the manufacturing phase represents a higher level of complexity of the AI project. At this stage, you do not need to prove the solution is effective or not, it must show that it can integrate with the company’s infrastructure and work well in real conditions.
“To achieve success, Machine Learning projects need to look closely at the structure and size of the company, its customers, and the internal workflow.”
It is the infrastructure, knowledge and data management that are the barriers that prevent the PoC project from reaching the next stage of production, and most businesses are still not aware of the importance. of the process of making PoC a real product. During this period, it is likely that the enterprise will have to change the existing working system. Besides, the closer to the final release stage, the more problems will arise.
What is the production phase? A system or solution is put to use in real life. It is no longer a PoC project to test whether the solution works, or test on sample data, which is the real data used to solve real problems.
After many projects, I realized that most AI solution providers could not prove the idea they originally came up with. But why is the transition from PoC to Production being considered a nightmare for AI projects? In fact, these projects do not reach the production stage because of the following factors:
- PoC did not produce the expected results
- Operating costs are too great
- Too complicated for the company
- Data is missing
- The parties concerned are not satisfied
A company may be willing to abandon an AI project if it faces data problems or a new workflow, even though they find that this AI solution is perfect for business. The company is having. In fact, organizations will have to deal with a range of software-related issues, data security and the amount of data required for training before entering the production phase. Another aspect is that companies often underestimate the cost needed to build an AI model that works on a real scale. Getting a prototype into production takes more investment than the business thinks, and managers need to make sure they can afford it when the project goes into production.
“The PoC project that applies your machine learning is just the first step in a long journey. You need to look beyond the issues that might arise when you decide to expand the project into a production system at scale. real.”
2. Why your AI PoC project failed
The PoC roadmap for an artificial intelligence project poses numerous challenges for businesses. From the lack of data, legal issues to the fear of AI-enabled applications or integration capabilities, businesses need to carefully analyze the various factors before putting the model into production. Companies should invest in PoC projects to learn about their own potential, improve data culture; at the same time, quickly finish projects that have no future and find the most potential approaches to continue investing resources. Many companies tried to make money with PoC first and chose a complex problem to solve through Machine Learning. And it was a one-way ticket to failure!
Enterprises should also understand that the skills needed to build a PoC are completely different from the skills to expand ideas for production. It sounds obvious, but without the structure to support the integration of Artificial Intelligence, the project, no matter how perfect, will fail. An AI project needs to be supported by management, and without long-term investment interest, the AI application will never reach any meaningful level of scale or usefulness. . Successful development of technology projects, especially artificial intelligence, takes time and patience.
In order for the PoC project to produce good results, businesses are required to conduct extensive research, build a multi-functional team that does a lot of work, and start searching and checking a wide range of specifications. hard. Companies can also consult with outside experts to refine the model. Although the team can prototype in just 2 to 3 weeks, the next steps will take longer and require a certain amount of money and time. Based on personal experience, implementing a good PoC usually takes 1 to 2 months. In particular, the data collection process is really time consuming. Many companies have a great idea of how to leverage AI but don’t have the right data to do it. For example, we have a PoC project that has an algorithm that demonstrates the ability to recognize faces captured in a condition of light, distance and angle. Thus, during the test, this algorithm needs to be trained in the changes of light, distance and angle of shooting, even skin color, gender and many more variants. In other words, algorithms need to be provided with more data to produce better results.
Understanding the difference between simply adding data to a PoC using the Machine Learning model and maintaining it on a large scale is very important. However, this aspect is often underestimated. After implementing a lot of projects with different data sets and mostly imperfect data, it can be concluded that people who try to convert desktop-only algorithms to production scale often have the tendency to downplay the importance of the time and energy required to convert machine learning algorithms into usable formats. It is important to minimize the distance between the actual requirements and the PoC data set. In this case, it is best to use the actual data itself.
Businesses must determine that it will take a long time to build a consistent and consistent data set. There are specific processes that require businesses to follow to produce data that meets the standards necessary to train a predictive model. When the PoC project is successful, many AI research teams believe that it is possible to prepare training data for the entire project. However, the reality is not so simple. They do not understand how difficult it is for the company to be able to produce the necessary data (work silos, slow organization …). At this stage, we begin to understand how the company works.
In fact, algorithmic training for additional practical cases is an integral part of the manufacturing process, and this training will create demand for a larger dataset.
On December 7, 2019, the rubikAI project with Up Co-working Space will hold an AI Series Talks # 2 in-depth discussion with the topic: “From AI PoC to reality: How to be successful? “ The event will take place from 14:45 – 17:00 at BK HUP, 17A Ta Quang Buu, Bach Khoa, Hai Ba Trung, Hanoi. Join the event at http://bit.ly/AI-Series-Talks-2 to not miss the opportunity to learn about the journey to build PoC and draw lessons for your own business!