Deep Learning

What is Deep Learning?
First, let's understand the basics of Artificial Intelligence (AI) and Machine Learning (ML):

Artificial Intelligence: Any technique that enables computers to mimic human behavior.
Machine Learning: The ability of a system to learn without being explicitly programmed.
Deep Learning: A subfield of machine learning that extracts patterns from data using neural networks.

In deep learning, we teach computers how to learn a task directly from raw data.

Why Deep Learning?
Manually engineering features is time-consuming, brittle, and hard to scale. Deep learning answers the question: Can we learn useful features directly from data?

Example: How Would You Detect a Face?
Imagine I tell you to build an AI that can detect faces in pictures. How would you even start?

First, you'd look for simple things like lines and edges.
Then, you'd detect curves — like the roundness of an eye or cheek.
Next, you'd combine those curves and lines to identify facial parts — eyes, nose, ears.
Finally, you'd assemble all those parts to recognize a full face.
This is how humans intuitively recognize patterns — from small pieces to the bigger picture.

Now here’s the powerful part: deep learning does this automatically. You just feed the system enough images, and it learns these steps on its own — layer by layer.

So What's the Big Idea?
The main idea of deep learning is this: You don’t need to hand-code every step. Just give the model enough examples, and it will figure out what patterns to look for — from simple lines to full faces.

That’s what makes deep learning so powerful.

Why Now?
Neural networks have existed for decades. So why is deep learning suddenly everywhere?

Because three major things have changed:

Data: We now have massive amounts of data from phones, social media, sensors, etc. Deep learning thrives on data.

Compute Power: GPUs are now fast, cheap, and widely available, making it possible to train large models efficiently.

Open Source Tools: Frameworks like TensorFlow, PyTorch, and Keras make it easy to build deep learning models without starting from scratch.

So while the ideas are old, today's tools, data, and hardware finally let them shine.

Source:course MIT 6.S191 MIT:Deeplearning

License: Creative Commons BY-NC-SA 4.0