“Those of us in machine learning are really good at doing well on a test set, but unfortunately deploying a system takes more than doing well on a test set.”, Dr. Andrew Ng, who co-founded Google Brain and Coursera, is currently CEO of DeepLearning.AI, said.
When you plan to develop AI solutions to solve business problems, you not only need to let AI and ML professionals find suitable machine learning models, train and test the models with data and use the models to predict information that you feel is valuable. You must also integrate AI systems to work seamlessly with your existing systems (e.g. inventory, CRM, ERP, accounting etc.) and many operational workflows.
We recently participated in a project that uses an artificial intelligence recognition system to identify road defects. This is a system that uses a supervised machine learning model. The training and test data are thousands of classified images of road defects, and the machine learning model will be used to predict the health of the road. If you are familiar with AI and ML, you might think that it does not seem too difficult. (By the way, preparing the classified images here may be the most time-consuming and demanding manpower resource.)
However, the complete project requires resources to make the artificial intelligence recognition system suitable for actual business use. First, in order to detect defects, cameras on a driving vehicle will be used to continuously take images along the road at regular intervals, or videos are recorded along the road and then a backend computational process is applied to convert the videos into useful images. The project will also cover the development of a front-end application for edge computing devices (mobile app and phones here) as well as cloud computing integration (uploading video or image files securely to the cloud storage area). The front-end application will be used by some frontline personnel on the roads.
In order to use the AI prediction results, an ERP system will be used to integrate relevant information and work processes. For example, upon receiving defect alarms and GPS (Global Positioning System) data from a specific location or area, ERP automatically sends out an inspection and repair order request to a dedicated supervisor or manager for approval of actions. This intelligent system provides user interfaces for related workflows, management and operation reports.
In short, deploying an AI system (for business use) takes more than doing well on a test (training) data set.
Image source : Wikimedia Commons
Suggested reading :