Build the mental model
Traditional machine learning depends on a human hand-engineering features: a spam filter's designer decides in advance that 'contains the word winner,' 'sent at 3am,' or 'has an unusual sender domain' might matter, then a simple model (like logistic regression) learns to weigh those hand-picked signals. This works only as well as the human's intuition about what matters, and breaks down for raw, unstructured data like pixels or audio waveforms, where no human can hand-craft the right features. Deep learning removes that bottleneck: a neural network is fed raw data directly and learns its own hierarchy of features through training — early layers might learn to detect edges in an image, middle layers combine edges into shapes, later layers combine shapes into objects, all discovered automatically from data rather than specified by a person. This isn't new math — the core ideas (layered networks, backpropagation) are decades old — what changed is that we now have enough labeled data and enough compute (especially GPUs, built for exactly the parallel matrix math networks need) to actually train networks with millions or billions of parameters until they work well.
Connect it to a real scenario
This course builds toward giving the Tutorial Platform real deep-learning-powered features: classifying whether learner feedback comments are positive or negative, ranking search results by a learned notion of relevance instead of fixed rules, and recommending each learner's next lesson using a model trained on past learning patterns. None of that is possible with hand-engineered rules alone — feedback text and learner behavior are far too varied for a person to enumerate every useful signal. Over this course you'll build the actual mechanics behind such features, starting from tensors and gradients in this chapter, through the network architectures (CNNs, RNNs, transformers) that later chapters use to actually solve these problems.
Try the working example
TRADITIONAL MACHINE LEARNING
-----------------------------
Raw Data --> [Human hand-engineers features] --> Simple Model --> Prediction
(email) "contains 'winner'?" (e.g. logistic (spam?)
"sent at 3am?" regression)
"odd sender domain?"
(human decides which signals matter)
DEEP LEARNING
-----------------------------
Raw Data --> [Neural Network learns its own features] --> Prediction
(pixels) Layer 1: edges
|
Layer 2: shapes
|
Layer 3: objects
(network discovers which signals matter, from data)The diagram shows that traditional ML pipes raw data through features a human chose before a simple model ever sees it, while deep learning feeds raw data straight into a network whose layers discover increasingly abstract features (edges, then shapes, then objects) on their own during training.5-minute try-it
Pick a real-world classification problem (e.g. detecting cats in photos) and list 5 features a traditional ML approach would need a human to hand-engineer — then note why each would be hard to define precisely by hand.
One important caution
Thinking deep learning is a brand-new mathematical breakthrough — the layered-network and backpropagation math is decades old; treating it as magic instead of understanding what changed (data + compute) leads to unrealistic expectations.
Assuming deep learning always beats traditional ML — for small, tabular datasets with well-understood features, a simple model (e.g. logistic regression or gradient-boosted trees) often performs just as well with far less data and compute.
Wikipedia — Deep learning — Deep Learning