Deep Learning
Deep Learning
What is deep learning?
- Multi-layer neural networks that learn features on their own
- A powerful branch of machine learning
- The engine of the modern AI revolution
What deep learning has really changed is not the accuracy numbers but the order in which a project runs. It used to be that domain experts spent more than half the schedule distilling "which values indicate a defect" into rules, and when those rules were off, the model was off with them. With deep learning, data collection and label quality control take that place. So budget lines shift from consulting fees to GPU usage fees and labeling labor costs, and the skill required of business-side staff moves from "putting rules into words" to "collecting usable data and defining the criteria."
Deep learning is not the answer to every problem. It dominates on unstructured data such as images, speech, and natural language, where features are hard to define in words. But on spreadsheet-style tabular data, such as customer churn prediction or credit scoring, traditional boosting-based machine learning is often faster and more accurate. Data volume matters too. Stack deep layers on a few hundred to a few thousand records and the model simply memorizes them (overfitting). In that case, transfer learning (fine-tuning), which takes a large pre-trained model and adjusts it slightly with your own data, is the realistic choice. Also weigh the costs separately: training is a large one-time expense, but inference costs recur every month.
Projects most often break down on the expectation that 'it’s deep learning, so it will figure things out by itself.' If labels are tangled, the model learns that tangle as is, and on data with a 2% defect rate, labeling everything 'normal' still yields 98% accuracy. A more common failure is the gap between the PoC environment and real site conditions. It is not unusual for an inspection model that scored 97% under lab lighting and cameras to drop to the 30% range after a single lighting change on the line. The model also cannot explain its decisions in words, so in areas where you must be able to explain them, such as loan screening or HR evaluation, write the retraining cycle and the human final-check procedure into the contract up front. Otherwise you will be left to deal with the fallout.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Ara LimExpert in global leadership, AI, and business communicationRooted in neural network research, it took off with work by Geoffrey Hinton and others in 2006, then with GPUs and big data in the 2010s.
Consider a hypothetical case. Company A, an auto-parts manufacturer with 320 employees, relied on 12 inspectors working three shifts for visual inspection. Each inspector applied different criteria, so the same product passed on Shift A and failed on Shift B, and customer complaints about defects reaching them kept coming at 4 per month. The company launched a six-month project to introduce deep-learning visual inspection.
- When Mr. Kim, head of the quality team, objected, “If we can’t state the criteria ourselves, how would AI know?”, the executive in charge proposed starting with the products where judgments disagreed. They secured 40,000 inspection images covering three months.
- Three experienced inspectors spent two weeks tagging 6,000 defect images into three types: scratch, dent, and foreign matter. The 800 images on which the three disagreed were pulled from training and sent to a separate criteria meeting.
- Instead of training from scratch, Mr. Park, the AI lead, fine-tuned a public pre-trained model on the company’s own data and confirmed 96.8% PoC accuracy in three days on just 2 GPUs.
- In the first week on the floor, missed detections suddenly rose, and the cause was a change of line lighting. Mr. Park reshot 5,000 images under the new lighting and retrained the model, then documented monthly retraining as an operating rule.
- Six months later, defect-escape claims fell from 4 a month to 0.5, and inspection time per product dropped from 8 seconds to 2. Seven of the 12 inspectors were reassigned to final review and equipment maintenance.