Artificial Intelligence

Types of engineers

Before you hire a Machine Learning Engineer hire,

• Data Analyst
• Data Scientist
• Data Engineer (everyone depends on them!)

Data Analyst

• What: Analyse prediction algorithms (models).
• Uses: Excel, SQL (Bonus: R / Python)
• Make direct analysis on data - less bugs and trying to get stuff to run.

Insight Analyst

Mainly: Look at data and extract useful data to later use.

Additional models

• Simple models like linear regression

• Time-series forecasting package from Facebook: FB Prophet

  • Spot weekly trends
  • Christmas trend (e.g. PosterXXL as Christmas gift)

Data Science

• What: Write prediction algorithms (models).
• Uses: Python, R
• Senior Data Scientist: Write Production-Code

Linear Models

Decision tree - Bäume mit Regeln

E.g. **Gradient Boosting** and **Random Forst**:  Thousands of such trees are agregated.

• Examples:

  • Theft detection:
  • Kündigungsrate vorhersagen

Natural language processing

Customer feedback form filtering: Sentiment Analysis

Neural Networks

• My talk about how similar Neural Networks are to the human brain

Deep Neural Network vs. Statistical Data Algorithms

Biggest gain (as compared to classical neural networks) with the following data: text, images, videos.

Simple algorithms vs. costs to maintain a neural network.

Input: # of rooms, pool?, garden?, living space in square meter Output: Price of house

Recurrent Neural Networks

Examples: -

• Composing music

• Get data into network in certain order (as compared to all at once in deep neural networks)

  • Given inserted words, deduce what the next word in the sentence may be.
  • Compose music

• Training of recurrent neural networks needs a lot of time and costs money

  • 200000 pounds costs to train network
• State-of-the-art

LSTM

• LSTMs forgets unimportant information.
• Standard RNNs possibly forget important information
• My talk about Recurrent Neural Networks (German)

Convolutional Neural Networks

A deep neural network where each layer is a

First layer

Erste Shicht: Simpel, nur Schatten.

• Augen und Nasen, dann Mund: Dunkle Punkte werden gesucht.

  • Nächste Schicht findet mehr details

Data Engineer

• Set up database
• Make sure data is available
• Data streaming

production code

Big data

Structured data

• Data lies in databases or so called data lakes.
• Basically, only Banks, Amazon or Facebook have BIG data.
• For example, 1 Mio lines in SQL table does not qualify as BIG data yet.

If own database crashes, Big Query from Google e.g. needs 2min to run through 2TB of data.

Unstructured data

• For example: CV

Examples

Retail

ClickStream

Recommender Systems

1. Generic: Pick top sellers and other values from descriptive statistics.

2. Content based filtering: Article about covid 19 -> articles about pandemics

   Similarity based

Spotify: Collaborative filtering

- user based attributes and product grading
- You like `foo`, other people similar to you like `bar`, so the algorithm will recommend `bar` to you.