Deep Learning is a type of machine learning based on artificial neural networks in which multiple layers of processing are used to extract progressively higher level features from data.
It is turning things (data) into numbers and finding patterns in those numbers.
| Traditional Programming | ML Algorithm |
|---|---|
| It is hard coded explicitly | The Algorithm figures out the code |
For complex problem. Imagine, can you hard code a self driving car? It would be difficult to do so for a dynamic environment. So, why not allow machine learning (deep learning) figure it out...
Machine Learning can be used for literally anything as long as you can convert it to numbers and program it to find patterns. Literally, it could be anything; any output or input from the universe.
If you can build a simple rule-based system that doesn't require machine learning, do that! - Google Machine Learning Handbook.
- Problems with long list of rules: when the traditional approach fails, machine learning/deep learning may help.
- Continually changing environment: deep learning can adapt (learn) to new scenerio.
- Discovering Insight within large collection of data
- When you need explainability: the pattern learning by a deep learning models are typically uninterpretable by a human.
- When traditional approach is a better option
- When errors are unacceptable: since the output of deep learning model aren't always predictable.
- When you don't have much data: deep learning models usually require a fairly large amount of data to produce great results.
Machine Learning |
Deep Learning |
|---|---|
| Used for structured data | Unstructured data such as natural language text, image recognition, web article, voice recognition. |
| Algorithms: Random Forest, Naive Bayes, Nearest Neigbour, Support Vector Machine etc..These are often referred to as Shallow Algorithm | Algorithms: Neural Networks, Fully Connected Neural Network, CNN, RNN, Transformer. |