> For the complete documentation index, see [llms.txt](https://dsic-alcoi.gitbook.io/the-new-revolution-of-generative-ai-models/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://dsic-alcoi.gitbook.io/the-new-revolution-of-generative-ai-models/basic-concepts-on-generative-ai/deep-learning.md).

# Deep Learning

Although based on known methods, 10 years ago the party started

Yes, it a a subset of ML, but believe us, it's the **main reason of current AI hype**

<figure><img src="https://4092288392-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FY8AeUFC5KYkaUv3lbFhX%2Fuploads%2FLr5X0YfaYMxpi13AHexy%2Fimage.png?alt=media&amp;token=9972208f-db4e-44c2-9f70-9a55d74b4151" alt=""><figcaption></figcaption></figure>

## Structured vs. Non-structured information

**Traditional ML** techniques work very well with what we call '**structured information**'

Think of structured information as a classical database table, with columns as attributes and rows as examples or singular elements.

However.... in front of **unstructured information traditional ML techniques struggle (images, signals, sounds, etc.)**

Even with structured information, someone has to select a collection of features (difficult to get, difficult to define and select etc.). An expert in the problem is required, and many trial an error experiments and evaluations.

Find a good collection of features for a problem is called **feature engineering**. To find an optimal set of features for a specific problem is difficult (e.g. what symptoms for detecting a disease?)

And, again, in front of **unstructured data** the problem becomes even harder.

**Deep Learning (DL)**, as a representation learning technique, will try:

* Learn from input samples the rules
* But, at the same time, **find and decide what are the features** the technique has to pay attention to in order to create the model, with no human interaction at all