01_getting-started.Rmd

# Getting started

**Learning objectives:**

- What do I need to get started?
- Motivating examples
- Big picture deep learning with fastai

## Tools and Resources{-}

- python
- pytorch
- fastai: built on pytorch
- jupyter notebook
- cloud servers
- Meta Learning book
- Deep Learning for Coders book (materials in videos presented differently)

## Prerequisites {-}

This course doesn't teach python.

**What you don't need to do deep learning**

Myth (don't need) | Truth
------------------| ------------------------------
Lots of math | Just high school math is sufficient
Lots of data | We've seen record-breaking results with <50 items of data
Lots of expensive computers | You can get what you need for state of the art work for free

## Is it a bird system? {-}

- reminder: images are numbers!
- What can be an image? sounds, time series, trajectory data
- introduce python workflow for getting images from duckduckgo
- `DataBlock()` for setting up data for CV models
- runs a CNN live while streaming on 400 images

## Where is deep learning? {-}

- course occurred ~ 2 years ago (July 2022)
- dall-e: text to images
- midjourney: text to images
- deep learning + art
- Google Pathways language model: answers questions w/ explanations

- ethics.fast.ai: another course

## Deep Learning {-}

- With a NN, we don't hand code features, but it learns them through examples
- Deep learning composes these features
- In general, the deeper features are more specialized/localized

## Jupyter and the cloud {-}

- literate programming: interweaves code cells and markdown cells
- colab
- kaggle: tutorial and course code available
- course instructor claims to use a lot of functional style

## DataBlock() {-}

- getting data to model is most important data science task
- most existing model architectures work well across systems
- main arguments:
    + what type of input?
    + what type of output?
    + where is the training data?
    + how should the data be split? (validation set is a requirement)
    + where is the training output?
    + how should the training items be transformed?
- then use a learner to pair the data with the model type
- intermediate, flexible approach
- specialized functions for common data structures

## Pretrained models {-}

- fastai is *fast* because it uses pretrained models
- pretrained: don't start with a random network, but with weights from a model previously trained on a different dataset
- used in image models

## Types of problems where DL is state-of-art {-}

- computer vision: identify what's in an image
- segmentation: combine pixels that correspond to one object/item/etc
- tabular analysis: regression/classification
- collaborative filtering - recommendation systems
- NLP
- Robotics
- Playing games
- Medicine/biology applications

## When does DL win? {-}

Is it something a (expert) human can do quickly? 

Does it take a logical thought process based on sparse data?

A trained model is just something that maps inputs to results like another computer program, but getting to the end model takes some tricks.

## Course Video HW {-}

- Experiment with week 1 lab
- Read book chapter

## Meeting Videos {-}

### Cohort 1 {-} 

`r knitr::include_url("https://www.youtube.com/embed/URL")`

<details>
<summary> Meeting chat log </summary>

```
LOG
```
</details>