One day, Party A of a factory came to you and said that the labor cost of a process link is too high, and see if it can be replaced by artificial intelligence. After you go to the field for inspection, you find that their on-site workers rely entirely on experience, and the rules are very vague. After many investigations, you finally determined the general rules, and after you were able to abstract this process link in the form of artificial intelligence, the project was finally approved.
Then you find that data collection is particularly difficult (without electronic recording equipment, you have to copy the data on site, and there is no historical data, and it takes a long time to copy it, otherwise the data is not enough to train the model), and the collected data is marked dirty (because Inexperienced, sometimes I have to ask on-site workers to label, because there are basically no statutory labeling rules, and the data is dirty).
After all the hard work to collect the data, the training model went very smoothly. Find a resnet/bert to train it, and the accuracy is 95%. Then you report to Party A excitedly. Party A says that we don’t have a machine with a graphics card, only an old CPU server, and let you try to deploy the model on it. Alright, let's compress the model. After a lot of tuning, the model is fast enough on the CPU server, but the accuracy can only reach 90%.
The user found it feasible after looking at it, and then put forward new requirements: the project will be reported after a while, can you explain the principle of your algorithm to the experts in the province; 90% accuracy is still too low, can you get 100% %; Our process may need to be changed, your model should be adaptive...
This is also the reason why many leading manufacturers are unwilling to go deep into the industrial field to do deep learning applications: either because of various factors, they cannot do it (such as project problems cannot be abstracted), or they can do it too cost-effectively (requires investment) A lot of manpower, the project funds are still small, and the needs of each factory are different and cannot be reused).