ML vs Deep Learning: The Difference in Plain Words

ML vs deep learning isn't a contest between two rival technologies. Deep learning is one kind of machine learning, the kind built from neural networks with many layers, so anything true of machine learning in general is also true of deep learning. The difference that actually matters comes down to a single question. Who decides what the computer looks at?

In classic machine learning, a person does. Someone picks the handful of measurements the model gets to see, and the model learns how much each one counts. In deep learning the network is handed the raw material, pixels, sound, text, and works out for itself which patterns are worth noticing. The 2015 review by LeCun, Bengio and Hinton says the layers of features in a deep network are not designed by human engineers but "learned from data using a general-purpose learning procedure."

Every other difference the comparison pages list, more data, more computing power, harder to explain, follows from that one. I'll show you why, and I'll show you the year it stopped being a theory and turned into a scoreboard.

Who Picks the Features

The clearest example I've found is in the opening pages of the Deep Learning textbook by Goodfellow, Bengio and Courville. A simple machine learning method called logistic regression can recommend whether a birth should be a cesarean delivery. It doesn't examine the patient. A doctor types in a few relevant facts, such as whether there's a uterine scar, and each of those facts is what the field calls a feature. The model learns how strongly each feature correlates with outcomes, and that's all it learns. In the book's words, "it cannot influence how features are defined in any way." Hand the same model an MRI scan instead of the doctor's report and it would be useless, because individual pixels have negligible correlation with anything that goes wrong in a delivery.

That's the whole boundary in one picture. Classic machine learning eats the doctor's report. Deep learning eats the scan.

The textbook draws this as four flowcharts, with the boxes that learn from data shaded. Here's the same figure as a table, with an everyday example bolted onto each row.

The middle column of the classic row is a real job with a name. Google's machine learning glossary calls it feature engineering and describes it as determining which features might be useful, then converting raw data into efficient versions of those features, with the example that "you might determine that temperature might be a useful feature." That sounds tidy. The 2015 review describes the same job as requiring "careful engineering and considerable domain expertise," and the textbook says that for a complex task, manually designing features "can take decades for an entire community of researchers."

The textbook's illustration of why is the one that made it click for me. Suppose you want a program that detects cars in photographs. Cars have wheels, so the presence of a wheel would be a useful feature. Now try to describe exactly what a wheel looks like in pixel values, with shadows across it, sun glaring off the rim, a fender hiding half of it. You can't. Nobody could. So instead of a person writing "wheel," a deep network is given millions of photos and learns something wheel-shaped in its middle layers because that turned out to help it tell cars from not-cars.

A made-up example, to show the same split on ordinary business data. Say a subscription company wants to predict who'll cancel next month. The classic route has an analyst choose columns, months as a customer, support tickets in the last 90 days, plan price, and a model learns weights for them. The deep route feeds in the raw recordings of every support call and lets the network find whatever in the audio predicts a cancellation. The second one might do better. The first one you can explain to your boss in a sentence, and it'll run on a laptop this afternoon. If I were the analyst, I'd ship the first and keep the second as an experiment.

The Year the Line Was Crossed

For image recognition there's a scoreboard, and it settles the question better than any definition. The ImageNet Large Scale Visual Recognition Challenge ran every year from 2010, asking programs to label about a million web photos into 1,000 categories. The organisers' own paper reports the winning method and its error rate each year. Top-5 error means the right label wasn't among the program's five guesses, so lower is better.

Read the third column with the first table in mind. The 2010 and 2011 winners are classic machine learning, world experts designing feature extractors by hand and feeding a classifier. A year of that effort moved the error by 2.4 points. The next year, with learned features, it moved by 9.4. That single jump, 2011 to 2012, is the one I'd circle. The review authors describe 2012 as "almost halving the error rates of the best competing approaches," and the challenge paper says that by 2014 "all of the top contestants were relying heavily on convolutional neural networks." Its summary of what changed is the sentence I'd have every beginner read twice. It became possible to learn networks "directly from the image data, without needing to create multi-stage hand-tuned pipelines of extracted features and discriminative classifiers."

There's a human row, too, and it comes with a caveat the paper is honest about. One expert annotator, after training on the categories, labeled 1,500 test images with 5.1% error, against 6.8% for GoogLeNet on the same sample. A second annotator who trained on only 100 images came in around 12.0%. Fine-grained breeds of dogs, birds and monkeys took the humans minutes per image. So "better than human" here means better than a hurried human, and still a little behind a well-drilled one.

How those networks actually see, edges first, then shapes, then parts, is the subject of the sibling piece on convolutional networks. This article only reports the score.

How Many Layers Make It Deep

Here's something the ranking pages state as fact that the primary sources don't agree on.

Four authorities, three different numbers and one refusal to pick. I'd take the refusal seriously. The textbook's fallback definition is that deep learning involves "a greater amount of composition of either learned functions or learned concepts than traditional machine learning does," which is the feature idea again, worn differently. What makes a network deep is that each layer builds on the one before, and the review describes exactly that stack for images. The first layer's features typically represent edges at particular orientations, the second spots motifs made of edges, the third assembles motifs into parts of objects, and later layers detect objects as combinations of parts. Counting layers is a proxy for that. If you want the layer itself pinned to numbers, the neural net walkthrough builds one by hand.

Why Deep Learning Instead of ML

Because the features are learned, deep learning needs more of everything the features would have supplied. More examples, since it has to discover from scratch what the analyst would have told it. More computing, since it's fitting millions of numbers instead of a dozen. And it's harder to explain, since the features it found don't have names.

That's the pattern behind every comparison table on the ranking pages, and it's enough to make the decision most beginners are actually facing.

The two data numbers are the textbook's, and the authors date them to 2016 themselves, so treat them as the order of magnitude rather than a threshold to hit. Working well with less data is, in their words, an important research area, and it has been ten years.

The Claim That Doesn't Hold Up

One of the pages ranking for this search, Coursera's guide, says that "machine learning models require human intervention when they get something wrong, whereas deep learning models can learn from their own mistakes." I understand what it's reaching for, but as written it's wrong in a way that will confuse you later.

Both kinds learn from labeled examples in the supervised setting that covers nearly everything you'll meet. The 2012 ImageNet network learned from about a million photos that people had labeled. The cesarean model learned from records where the outcome was known. Neither one corrects itself once it's deployed. If a deep network keeps mislabeling your cat as a fox, no amount of it "learning from its own mistakes" fixes that. Someone gathers more labeled cat photos and trains again.

The human step deep learning removes is the feature-design step, and only that one. Humans still label the data, choose the network's shape, pick the training settings and decide when it's good enough. Keep the correction step and the feature step separate in your head and most of the fog around this topic clears.

The Syracuse page at the top of the results has its rows pointed the right way, more data, more compute, less explainable, but gives no source and no number for any of them. That's why the tables above cite where each number came from.

Is ChatGPT Machine Learning or Deep Learning?

Both, with no contradiction. The model behind the first version, GPT-3, is described in its own paper as "an autoregressive language model with 175 billion parameters." Those parameters are the learned numbers in a very deep neural network, which makes it deep learning, which makes it machine learning, which makes it AI. The nesting is the entire subject of the AI versus machine learning explainer, so I won't redraw the boxes here. What matters for this page, to my mind, is that nobody hand-designed features for it. It was fed raw text.

The Three Kinds of Machine Learning

People also ask about the three types of machine learning, and they're a different axis entirely, which is worth saying because the two axes get tangled. Supervised, unsupervised and reinforcement learning describe what kind of examples the machine is given. Classic versus deep describes whether a person designed the features. The ImageNet winners are supervised and deep. The cesarean model is supervised and classic. You can have any combination, and the plain guide to what machine learning is walks through the three types with kitchen-table examples.

Can You Learn ML in Three Months?

Three months is enough to understand classic machine learning properly and to run one guided model end to end, in my judgment, provided you spend it on the concepts in this article rather than on collecting frameworks. Feature, label, training set, error rate. Those four words carried every table above.

Deep learning is a second three months, after that, and it goes far better once the classic version isn't a mystery. Which course to spend those first months on, without a coding background, is the whole of the machine learning course picks piece. If you only take one thing from this page, take the question. Whenever someone says deep learning, ask who picked the features. The answer is the difference.