logo

ML Geo Curriculum
  • Machine Learning in the Geosciences

About this Book

  • Geosmart website
  • Acknowlegments
  • Primer

Chapter 1 - Open Source Ecosystem with Python

  • Getting Started
  • 1.1 Open Reproducible Science
  • 1.3 Jupyter Environment
  • 1.3 Python Ecosystem
  • 1.4 Computing Environments
  • 1.5 Version Control & GitHub
  • 1.6 Data Gallery

Chapter 2 - Data Manipulation

  • 2.1 Data Definitions
  • 2.2 Data Formats
  • 2.3 Pandas
  • 2.4 DataFrame Exploration
  • 2.5 Data Arrays
  • 2.6 Resampling Methods
  • 2.7 Statistical Considerations for geoscientific Data and Noise
  • 2.8 Spectral Transforms
  • 2.9 Filtering Data
  • 2.10 Synthetic noise
  • 2.11 Feature Engineering
  • 2.12 Dimensionality Reduction
  • 2.13 ML-ready data
  • Assignment: Preparing AI-Ready Data for The Final Project

Chapter 3 - Machine Learning

  • 3.1 Concepts in training supervision
  • 3.2 Classification and Regression
  • 3.3 Clustering: Unsupervied Classification
  • 3.4 Binary classification
  • 3.5 Multiclass Classification
  • 3.6 Logistic regression
  • 3.7 Random Forests
  • 3.8 Robust Training
  • 3.9 Ensemble learning
  • 3.10 AutoML
  • 3.20 Final Project - Classic Machine Learning
  • Homework Classic Machine Learning (50 points)

Chapter 4 - Deep Learning

  • 4.0 The Perceptron
  • 4.1 Neural Networks
  • 4.2 Multi Layer Perceptrons
  • 4.3 Model Training
  • 4.2 Physics-Informed Neural Networks
  • 4.3 Convolutional Neural Networks
  • 4.4 Recurrent Neural Networks: Processing sequences
  • 4.5 Auto-encoders
  • 4.6 NAS: Network Architecture Search

Chapter 5 - Workflow Management and Reproducibility

  • This chapter focuces on model workflow and ML reproducibility

Chapter 6- Introduction to Cloud Computing

  • Browser Access to Cloud Instances
  • Terraform Access to Cloud Instances
  • Cloud Provider ML Jupyterhubs

Reference

  • Glossaries
  • Bibliography
Powered by Jupyter Book
  • Binder
  • repository
  • open issue
  • suggest edit
  • .ipynb

4.6 NAS: Network Architecture Search

4.6 NAS: Network Architecture Search#

The best model is one that fits the data best and can generalize the best to new examples. Provided these two metrics, ML practitioners experiment on model architectures that best satisfy these conditions. This can lead to a time-consuming effort to find the best model architecture.

A more systematic approach to finding the optimal model is to perform a Network Architecture Search, which is basically a consistent method to explore the “model space”, as in the “model architecture space”.

The field of NAS yields several packages to automate the searches.

In keras, this is the “https://keras.io/keras_tuner/” or “auto keras”

For pytorch, this is auto-pytorch.

previous

4.5 Auto-encoders

next

This chapter focuces on model workflow and ML reproducibility

By eScience Institute, University of Washington
© Copyright 2022.