Why Do We Need Deep Learning?
Machine Learning can solve many problems, but some problems contain extremely complex patterns that are difficult to describe using simple rules or manually selected features. Deep Learning is designed to learn these complex patterns using neural networks with multiple layers.
The simple idea
Deep Learning is useful when the data is complex and the patterns are difficult to describe manually. Neural networks can learn these patterns from examples.
Why Do Simple Rules Become Difficult?
Imagine that you want a computer to recognize whether an image contains a cat.
You could try to write rules such as:
A Rule-Based Approach
The problem is that cats can look very different. They can have different colors, positions, sizes, lighting conditions, and backgrounds.
Writing rules for every possible situation quickly becomes impractical.
Example: Image Recognition
Deep Learning can learn useful visual patterns from many example images.
Instead of manually describing every possible cat, we provide examples and allow the neural network to learn useful patterns from them.
Why Are Multiple Layers Useful?
Deep Learning uses neural networks containing multiple layers. Different layers can learn different levels of patterns.
Earlier layers can learn simple patterns such as edges and shapes. Deeper layers can combine those patterns to recognize more complex structures.
Example: Speech Recognition
Speech is another type of data that contains complex patterns. A computer receives an audio signal rather than a written sentence.
Deep Learning can learn relationships between the audio patterns and the words that were spoken.
Example: Understanding Text
Words can have different meanings depending on the surrounding context.
Example 1:
I deposited money in the bank.
Here, "bank" refers to a financial institution.
Example 2:
I sat near the bank of the river.
Here, "bank" refers to the side of a river.
Understanding this kind of context requires learning relationships between words and their surrounding information.
Deep Learning Handles Complex Data
Traditional Machine Learning works extremely well for many structured problems. But some data contains much more complicated patterns.
Deep Learning is particularly useful when the model needs to learn complex relationships from these types of data.
Why Does More Data Help?
A Deep Learning model can benefit from seeing many different examples of the problem it is trying to solve.
Example: Cat Recognition
More varied examples can help the model learn patterns that generalize to new images.
However, more data does not automatically guarantee a better model. The data still needs to be relevant and useful.
Do We Need Deep Learning for Everything?
No. This is an important point.
Deep Learning is powerful, but using a more complicated model does not automatically make a solution better.
Example: Customer Churn
Suppose your dataset contains:
A traditional Machine Learning model may be enough for this type of structured data.
You should choose Deep Learning when its ability to learn complex patterns provides a real advantage.
When Is Deep Learning Useful?
Deep Learning is particularly useful for problems involving images, speech, video, language, and other complex data where manually defining all useful patterns is difficult.
Machine Learning vs Deep Learning
Remember this
We do not use Deep Learning simply because it is more advanced. We use it when neural networks can provide an advantage for learning complex patterns from data.
Check Your Understanding
Why do we need Deep Learning?
Some problems contain complex patterns that are
difficult to describe manually.
Why are multiple layers useful?
Different layers can learn increasingly complex
patterns from the input.
Does every problem require Deep Learning?
No. Traditional Machine Learning can be a better
choice for many simple structured problems.
Give two examples where Deep Learning is useful.
Image recognition and speech recognition are two
examples.