Machine Learning · Deep Learning
Machine Learning / Deep Learning
What is machine learning?
- Learning rules from data
- Deep learning is its most powerful branch
- Let machines learn the rules instead of writing them
The impact of adoption shows up first in how the work is sequenced, not in the code. In rule-based systems, developers added another conditional every time a new exception appeared. With machine learning, the default response is to collect the failed cases and retrain the model. That is why a large share of project effort goes into data collection, labeling, and validation rather than the model itself. If the business side cannot define what a correct answer is, even the best algorithm has nothing to learn from. This is why AI projects belong to the business units, not just the IT department.
Learning methods fall into three broad types. Supervised learning teaches with labeled answers (defect detection, churn prediction). Unsupervised learning finds groups without answers (customer clustering, anomaly detection). Reinforcement learning rewards trial and error (inventory placement, route optimization). What separates deep learning from the rest is who extracts the features. Traditional machine learning relies on variables that people design, such as “thickness,” “day of week,” or “time on site,” while deep learning picks up features on its own from raw inputs like pixels or characters. The trade-off is that it needs a lot of data and compute, so for a few thousand rows of tabular internal data, deep learning is not always the better choice. In that range, decision-tree methods are faster and easier to explain.
The most common slip in practice is the accuracy number. On a line with a 3% defect rate, a model that labels everything as good still scores 97% accuracy. Teams that cut visual inspection based on that number alone have ended up facing customer claims. Ask for detection rate and false-alarm rate to be reported separately. Another is the belief that a model is built once and done. When raw materials, equipment, or the customer mix change, the learned patterns no longer fit, performance quietly degrades (data drift), and months pass without anyone noticing. Unless you fix the retraining cycle and an owner in your operating rules up front, you end up with a model and no one accountable for it.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Jaryong KooMarketing PhD and leading authority on AI, data analytics, and personal brandingThe term “machine learning” is believed to have been coined in 1959 by Arthur Samuel of IBM. Deep learning rose in the early 2010s, as growing data and computing power led to major advances in image and speech recognition.
Consider a hypothetical case: Company A, an automotive injection-molded parts maker with 260 employees and annual revenue of about KRW 48 billion (≈ US$35,912,000). Nine inspectors checked appearance by eye, and on the night shift defect escapes kept recurring at around 30 a month, with customer complaints rising to 4 per quarter. The quality assurance team set up a 5-month project to introduce deep-learning vision inspection.
- The first thing Deputy Manager Park of quality assurance did was not write rules but collect photos. Using the line cameras, the team re-shot and stored 12,000 good-part images and 1,400 defect images from three months of production.
- Labeling hit a snag. When three inspectors tagged the same photo differently, some as “scratch” and some as “foreign matter,” Manager Park put 40 limit-sample photos on the wall and set one standard: “anything darker than this is a defect.”
- The first model scored 96% accuracy, but the production team was unimpressed. Defects made up only 3% of the data, so a model that labeled everything “good” would score 97%. The actual defect detection rate was just 62%.
- The production manager put it this way: “Better to over-reject than to miss one.” So the team lowered the decision threshold and sent every suspect item to a re-inspection line. Over-rejection went up, but inspectors only had to re-check 180 images a day.
- Six weeks later, a change of raw-material supplier altered the surface gloss, and the detection rate dropped to 71%. After that, monthly retraining on 500 new images was written into the operating rules, with Manager Park named as the owner.
- Five months later, defect escapes fell from 30 a month to 4, and customer complaints went from 4 per quarter to zero. Four of the nine inspectors were reassigned to re-inspection and equipment checks.