Course Content
Deep Learning Essentials
13 sections · 61 lessons
Why do we need Deep Learning when Machine Learning already exists?
What you need to know
The feature-engineering bottleneck
A classical model such as logistic regression, a random forest or XGBoost takes a table of numbers. Someone has to decide what those numbers are. For a loan model that is easy: income, age, number of late payments. For a photo it is not. A 224×224 colour image is 150,528 pixel values, and no single pixel means "leaf blight" or "cat".
Before deep learning, vision engineers wrote feature extractors by hand: edge detectors (Sobel), histograms of gradients (HOG), SIFT keypoints. These worked for narrow tasks and broke when lighting, angle or background changed.
What deep learning changes
A deep network stacks layers. Each layer turns its input into a slightly more useful description. In a vision model, the first layer reacts to edges and colour blobs, middle layers to textures and shapes, and late layers to object parts. Nobody programs these; they appear because they help reduce the loss.
The second difference is scaling. A classical model on tabular data usually plateaus: after some point more rows barely help. Large neural networks keep improving as you add data, parameters and compute, which is why the biggest gains of the last decade came from them.
When classical ML still wins
| Situation | Better first choice | Why |
|---|---|---|
| 20,000 rows of tabular data | Gradient boosting | Strong accuracy, fast, little tuning |
| Need to explain each decision to a regulator | Linear model or trees | Easy to explain |
| Images, audio, free text | Deep learning | Features cannot be hand-written |
| Millions of labelled examples | Deep learning | It keeps improving with scale |
Deep learning also costs more: GPUs, longer training, more labelled data, and harder debugging.
A real-life example
An agri-tech startup wants farmers to photograph a tomato leaf and get a diagnosis: healthy, early blight, late blight or leaf mould.
With classical ML, an engineer extracts colour histograms and texture statistics, then trains a random forest. It reaches about 70% on test photos, and drops badly on photos taken in shade or with soil in the background, because the hand-made colour features confuse brown soil with brown lesions.
With deep learning, the team fine-tunes a pretrained ResNet on 8,000 labelled leaf photos. The network learns lesion shapes and edge patterns that no one described to it, and reaches about 94% on held-out photos from new farms.
The same startup also predicts which farmers will repay a micro-loan. That data is a 30-column table of land size, crop type and past repayments. There, XGBoost beats a neural network and the credit team can read its feature importances. One company, two problems, two different right answers.
Follow-up questions to expect
- "Is deep learning a subset of machine learning?" — Yes. AI contains machine learning, and deep learning is the part of machine learning that uses multi-layer neural networks to learn representations.
- "Why did deep learning only take off around 2012?" — Three things arrived together: large labelled datasets like ImageNet, GPUs fast enough to train big networks, and training tricks such as ReLU, better initialisation and dropout.
- "What would you use for 5,000 rows of customer data?" — Gradient boosting as the baseline. I would only try a neural network if there were also text or images, or if the baseline clearly left accuracy on the table.