ML – «Programming the Unprogramable»

1. Introduction

AI/ML is not a futuristic topic anymore. For example the company Icons8 use AI/ML to generate headshots. This put stock photo companies on notice (here).

Developers and BI analysts can use AI/ML tools now, it is not only for Data Scientists.

This blog post shows ML.NET, which is a collection of cross-platform libraries and CLI to create and train ML models.

This post is a summary of the following video and this tutorial.

2. Important Concepts

• ML helps to solve challenging programming questions.
• Example of challenging programming questions:
  • Write a function which takes an image and tells if there is a face on it.
  • Write  a function to tell the price of the shirt based on textual description?
  • Write a function try to play rock paper scissors with a machine (think about different skin tones, different hand sizes, etc).
• Machine Learning creates a function (called Model) using sample data. In Machine Learning, you develop/train you custom ML models to infuse custom machine learning into their applications.
  • A model is the algorithm
• Interpreting the data is tricky. What do you do if you’re told your 82% healthy
  A more comprehensive answer would be, «here the characteristics you can improve, bla bla»
• There are Supervised and non-supervised training.
  • In supervised learning you use algorithms:
    • Regressions.
    • Clusters.
    • Classification.
    • Classification.
  • Machine Learning is: unsupervised training
• There are tons of out of the box ML Trainers (described in the Tooling section below).

3. Machine Leaning Stages

There is a training VS running

4. Working Samples for Developers

Clone the ML.NET sample git repo (link here). Some facts about the code:

• • About Data Inputs:
    • In this example there are two datasets:
      • One to train your model (called wikipedia-detox-250-line-data.tsv, which is larger) and test data (called wikipedia-detox-250-line-test.tsv) to determine how well model is performing.
    • There is also an input class called SentimentIssue to model the data in file: wikipedia-detox-250-line-test.tsv
  • About Data Processing:
    • Then you need to transform data (a.k.a. Featurize Text) for ML algorithms to understand it.
      • In AI/ML, arrays are the way how Images, text, etc are represented (more on that here).
      • 

• • • Last but not least, you need to apply ML algorithms ML (a.k.a. Trainers, see here).
    • 
    • How can you tell which one perform best for your scenario? -> ML.NET offers Model Builder can do it for you (more on that here).
    • 
  • Time for Predictions:
    • The whole point about building a model is to use it for Predictions.
    • Last but not least, you evaluate the accuracy of the model.
    • For the record, a model is a zip file. The zip file is what you deploy to PROD.

5, ML Tools

There are lots of ML/AI tools, in this post we provide below only a general introduction to ML.NET and BigQuery ML.

5.a ML.NET

There is a complete post about ML.NET [here], but in general with ML.NET you can do the following general tasks:

• Automated Model Generation
  •  Classification
    • Binary Classification
    • Sentiment Analysis
    • Spam detection
    • Credit Card Fraud Detection
    • Hearth Disease Prediction
  • Multi-class classification
    • GitHub Labelers
    • Iris Flowers Classification
    • MNIST
    • Support Text Classification
  • Models for Regression (to answer questions like how much? how many?)
    • Price Prediction
    • Sales Forecasting
    • Demand Prediction
• Recommendation
  • Product Recommendation
  • Movie Recommendation (Matrix Factorization)
  • Movie Recommendation (Field Aware Factorization Machines)
• Clustering
  • Customer Segmentation
  • IRIS Flowers Clustering
• Time Series Forecasting
• Ranking, etc.
  • Rank Search Engine Results
• Anomaly Detection
  • Sales Spike Detection
  • Power Anomaly Detection
  • Credit Card Fraud Detection
• Computer Vision
  • Training
    • High Level API
    • Featuizer Estimator
    • Predictions (Pretrained TensorFlow model scoring)
  • Object Detection
• Cross Cutting Scenarios

5.b BigQuery ML

In BigQuery ML, only two steps are needed to create a model and make predictions based on it, the syntax is inspired in SQL.

Thanks,

Javier Caceres