MODERN ERA • HISTORY OF AUTOMATION
Machine Learning: Weights Fit to Examples
Weights fit to examples. Not a student, and not a mind that picked up a trade
The last page was a guess about a machine, made from a file of its past. This page is how that kind of guess gets its numbers. You choose a task and a score. You hand over examples. A procedure nudges a pile of weights until the score on those examples improves. Later, the same weights are applied to a case they were not shown. That is the whole trick. Arthur Samuel, writing in 1959 about a checkers program, called it machine learning. His word stuck. His program did not understand checkers. It got better at a game it was scored on.
A cam, a relay, a PLC, and a taught robot path are still how a plant stops, permits, and moves. A model does not replace them. It is what you reach for when you cannot write the rule for every crack, and you are willing to be wrong in the way the examples were wrong. If a title says the machines learned, the title is using Samuel’s word for a life the file does not have.
A rule you can read, and a weight you cannot
If the sensor is on, stop. If the pressure crosses the line, open the valve. If the robot has permission, run the path. Those are the PLC page. A person can point at the rung. Some jobs do not offer you a rung. Every crack in a casting is not a sentence you will finish writing. Every glare on a wet part is not a threshold you will keep true. A model can be fit to examples of the crack and the glare. It can then miss a crack that was not in the file, and it can call the glare a crack because the file taught it to.
Use it where the rule book failed and the misses are tolerable or checked. Do not use it where the miss is a hand. The predictive page already said the safety trip stays in the controller. It stays there.
What the procedure actually does
You gather examples tied to the task. Inspection wants images, and someone has to have marked which ones were bad. A maintenance model wants the tags the last page already distrusted, plus a failure log that often says “other.” The procedure adjusts parameters to shrink a measure of error on that file. Then you try it on examples you held back. The held-back set is only a comfort if it is the plant. A test set drawn from the same week, the same light, and the same camera is a rehearsal. The Monday the supplier changes the resin is the test that counts, and it does not happen in the lab.
Memorizing the file is a failure mode with a polite name, overfitting. The score on the training examples looks excellent. The next lot looks like nothing it stored. The predictive page said a model can memorize a lying sensor. Same disease, any task.
SHOP LANGUAGE
Collect. Fit the weights. Try what you held back. Deploy only where a refusal still exists. Watch
Deploy is not “turn it loose.” A divert, a work order, or a robot move still needs a hard limit or a person who can throw the answer out. Monitoring is not a slogan. The plant drifts. The weights do not, until you fit them again.
One calculation that impressed a meeting is not a lifecycle. The lifecycle is the dull part, and it is the part that keeps the scrap honest.
Labeled examples teach the label, including the bad one
Supervised learning means the file contains an answer you supplied. Thousands of images marked good or bad. The procedure hunts for whatever in the image predicts that mark. If the mark was the inspector’s habit — a shadow counted as a scratch, a real scratch on the edge never marked — the model inherits the habit and applies it faster than the inspector could. Classification, a numeric estimate, a forecast of a tag: same bargain. The answer is only as good as the people who labeled, and the cases they never labeled do not get a vote.
No label means no answer, even when the clusters look neat
Unsupervised methods look for structure when you did not write the answer down. Groups of similar rows. A point that sits far from the crowd. A shorter set of numbers that still carries most of the variation. In a plant this is what you have when the fault log is thin, which is often. A cluster is not a failure mode. Unusual is not broken. The predictive page already said an anomaly flag is a request to look. Do not let a tighter algorithm promote the flag into a cause.
A reward is a score you wrote. Do not let it practice on the press
Reinforcement learning sets an agent in a loop: act, receive a number, adjust. Over many tries it hunts a policy that runs that number up. Richard Sutton and Andrew Barto’s textbook is the map of that idea. The number is not the good of the plant. It is whatever you coded — cycle time, a game point, a simulated success. If the code forgot a smash, a scrape, or a person in the aisle, the policy will spend those for free.
That is why the honest industrial use is a simulation, or a box so tight the live cell cannot be the practice field. The humanoid page already refused a policy that misses and calls it understanding. A robot that “learns on the job” by trying grasps on live product is a scrap program with a research abstract. Say so before someone turns it on.
The file is the raw material, and it is usually dirty
The procedure cannot see the bearing. It can see the channels you recorded. A bad calibration, one lighting setup, a failure text that changes with the shift, a camera moved an inch: the weights will fit those accidents. The predictive page walked through bad tags. I will not walk them again except to say the duty is stricter here, because a vision model fails in pictures and a maintenance model fails in dates, and both failures look like confidence.
More data is not automatically better data. More of the same week is a thicker copy of the same blind spot.
WHAT YOU CANNOT SKIP
The instrument, the label, the condition, and a test that is not the training week
A fancier network will not confess that the labels were sloppy. It will fit them more tightly. The model is one component. The refusal path is the other, and it is the one that keeps a bad score from becoming a moved part.
The “neural” story is a metaphor that got away
An artificial neural network is a stack of weighted sums and simple bends, arranged in layers. Training moves the weights. Frank Rosenblatt’s perceptron, described in 1958, was an early machine of that family, and it was real hardware. Marvin Minsky and Seymour Papert, in 1969, showed what a single layer of that kind cannot separate. That was a limit on a class of models. It was not the death of a science and not the birth of a brain. Later networks with more layers got past that particular limit. They did not become neurons.
Ian Goodfellow, Yoshua Bengio, and Aaron Courville say the biological resemblance is inspiration, and a loose one. Keep it loose. A weight is not a synapse. A layer is not a cortex. The computers page refused a mind in a stored program. Refuse it again in a picture of a brain drawn on a whiteboard.
Deeper stacks changed a photo contest. They did not change the interlock
Deep learning, in the narrow sense, means more of those layers, fit with a lot of examples and a lot of arithmetic. In 2012 Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton entered a deep convolutional network in the ImageNet photo contest and beat the prior entries by a wide margin. That is a real mark in the history of benchmarks. It is a contest of labeled photographs. It is not the morning a factory camera understood a casting, and it is not a date you should carve over the quality gate.
Speech, text, and pictures got better on their public tests in the years after, for the same family of reasons: data, machines that could multiply fast, and tricks in the training. Better on a test is a sentence you must finish. Better at what. On whose pictures. Checked by whom, on the line, this month.
A vision score is not yet a divert
Older industrial vision measured an edge, a blob, a dimension, with a rule an engineer could point at. A learned model takes examples instead, which helps when the defect will not sit still for a caliper. The warehouse page said a camera is a sensor with a lot of bits. The output here is a class or a score. A conveyor diverts only if some further rule, tested on purpose, says that score is allowed to move a part. “The prediction becomes the decision” skips the only step that makes the decision yours.
Light changes. The lens clouds. A new color of box arrives. The score keeps printing. Someone has to have decided, before that morning, what the line does when the score is shy.
INSPECTION, SAID IN ORDER
A picture. A score. A rule that may refuse the score. Then, maybe, a divert
Accept, reject, send to a person. The person is not a leftover from before the model. The person is the path for the image the file never contained. Take that path out and you have automated the confident miss.
Maintenance and robots do not get a new exemption
The predictive page holds the bearing. A learning method can combine channels. It can also date a death the physics did not announce. Do not reopen that argument here as if a deeper network settled it.
A robot path is still taught points until a model is allowed to propose a grasp or a detour. The proposal is a guess from examples of objects that looked like these. The humanoid page said a policy can complete the gesture on the wrong object and log success. Vision on an arm is the same risk with a wrist. Outside a fixtured cell, the misses get larger, not smaller, because the world gets larger. That is not the hour the robot learned the room.
Fluent language is not a procedure
Models that continue text can answer a question in sentences a technician recognizes. They can also invent a torque, a pinout, or a step that was never in the manual, in the same calm voice. A search of the real document is not the same tool as a paragraph that sounds like the document. If an engineer uses one to draft, a person who can be wrong in public still owns the draft. A robot instruction that arrived as a sentence is not safer because the sentence was grammatical.
I will not call that a simpler future between people and machines. I will call it a new way to be fluently wrong, useful only where the cost of the invention is caught before the tool turns.
The model proposes. The interlock disposes
When a score moves a cylinder, the miss is no longer on a screen. A bad class sends a good part to the scrap bin, or a cracked one to the customer. An alert that is usually false teaches the crew to mute it, which the predictive page already named. A trajectory from a vision model can be the wrong trajectory at full speed. Reliability here is not a virtue of the algorithm. It is the fence around it: what it is allowed to command, what it cannot command, and what happens when the score arrives late or not at all.
A safety function does not take a probability and hope. Guards, interlocks, a safety-rated controller: those stay deterministic on purpose. The model lives beside them. NIST’s AI Risk Management Framework, version 1.0 in 2023, is a way to talk about that risk. A framework is not evidence that a given model was fenced. Paper does not stop the cylinder.
The file ages. The weights do not notice
Wear, a new supplier, a redesigned part, a replaced sensor, a different shift’s lighting. The relationship you fit can leave. People call the leaving drift, and they argue about which kind. The practical sentence is shorter. The world moved. The model did not, until you measured the damage and fit again, or took it out. One successful trial is a week. The predictive page’s dull shift is the standard here too.
The person who knows the machine is not a bias to be removed
A model can rank patterns a busy crew will not plot. It cannot tell you the duct was built wrong, or that this scratch is the one the customer will actually return. An engineer, a technician, and an operator know different parts of that. The stronger system is the argument between them and the score, which is what the predictive page said about the wrench. Do not “automate away” the argument and then act surprised at the lot that looked nothing like the training week.
NO LADDER
A cam is not a program. A program is not a weight. A weight is not a mind
Geometry held an old motion. A PLC holds a rule you can read. A model holds a fit to a file. Those are three ways to store a behavior. They are not three steps of one creature learning to know.
“The way machines receive knowledge has evolved” is the sentence to drop. Knowledge is the word that smuggles the student back in.
A Jacquard card is not a learning algorithm
The loom stored a pattern for which threads rise. Change the cards, change the weave. The weaver stayed. The computers page already refused the bloodline from those cards to software. Refuse it again here. A card is not a weight. A weight is not a card that went to school.
Information can change what a machine does. That family resemblance is real. Descent is the part that is not.
What I will not promise, because this series stops here
I will not say vision will soon need no setup, or that robots will handle whatever box arrives, or that schedules will optimize themselves, or that a humanoid will be safe because a model was large. Those are wishes. A wish is not a result. What you can buy is a score on a task you defined, inside limits you built, checked by people who can overrule it. When the score is wrong in a way the file allowed, the plant pays. The home page of this site is a timeline of machines. It is not a promise that the last machine understood the first.
What I want you to keep
Samuel, 1959, used the phrase for a checkers program that improved a score. Mitchell’s definition is a task, a measure, and a file — not a mind. Supervised learning copies the label, including the bad label. A cluster is not a cause. A reward is a number you wrote; do not collect it on a live press. Rosenblatt’s perceptron was a machine, and Minsky and Papert’s limit was a limit, not a funeral and not a brain. AlexNet in 2012 won a photo contest. A contest is not a quality gate. A vision score diverts only through a rule you still own. Fluent text can invent a step. The interlock is not a probability. The file goes stale.
And the textile page was right. The cards held a weave. They did not teach anything to learn.
References & Further Reading
- Arthur L. Samuel, “Some Studies in Machine Learning Using the Game of Checkers,” IBM Journal of Research and Development 3, no. 3 (July 1959): 210–229. The phrase, and a program that improved at a game. Not a claim that the program understood the game.
- Tom M. Mitchell, Machine Learning (New York: McGraw-Hill, 1997). The working definition: a task, a performance measure, and experience. Experience in that definition is data, not a life. Author’s page.
- Frank Rosenblatt, “The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain,” Psychological Review 65, no. 6 (1958): 386–408. An early learning machine. The word “brain” in the title is the author’s frame, not a finding that the device had one.
- Marvin Minsky and Seymour Papert, Perceptrons (Cambridge, MA: MIT Press, 1969). Limits of a class of single-layer models. A limit, not the end of the subject, and not proof of the later layered networks.
- Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction, 2nd ed. (Cambridge, MA: MIT Press, 2018). The reward is whatever the designer encoded. Read it before anyone lets a policy practice on a live cell.
- Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” Advances in Neural Information Processing Systems 25 (2012). A photo-contest result. Not a factory acceptance test.
- Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning (Cambridge, MA: MIT Press, 2016). The textbook. The biological story is treated as loose inspiration. The book, free online.
- Christopher M. Bishop, Pattern Recognition and Machine Learning (New York: Springer, 2006), and Kevin P. Murphy, Machine Learning: A Probabilistic Perspective (Cambridge, MA: MIT Press, 2012). The math of fitting and of being wrong. Not a deployment record.
- Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th ed. (Hoboken, NJ: Pearson, 2021). A broad textbook. Already refused, on the humanoid page, as a verdict that a body understands. The refusal stands.
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), 2023. A way to talk about risk. Not evidence that any particular model on a line was fenced. NIST.
Nothing here treats a trained model as a mind or as a machine that learned a trade. A page title that says machines learned is a title, not a finding. The weights fit the score someone chose. The interlock stays outside the model.