Release notes AutoML v.0.3.1: score graph, solution README, and data science updates

Claudio Bruderer

Written by:
Claudio Bruderer (Product Manager at Modulos)

The latest release of Modulos AutoML is here. We’re shipping a number of improvements and new features designed to make your AI Journey better.

Data science updates

New Models

In v0.3.1, we’ve added a new model type: a CNN which takes as input a combination of an image and a table. This new model type opens up a whole range of new applications of AutoML where the data set consists of images with relevant metadata. An example of this would be satellite image data combined with coordinates and other properties of the location in the image.

New Objectives

When you train an ML model, you’re looking to maximize (or minimize) a score which tells you whether the model is doing well. In this release, we are adding two new objectives:

  • F1 (macro) score : evaluates the performance of a classifier like the Accuracy, but it gives smaller classes more weight and therefore it is useful for unbalanced datasets.
  • Median Absolute Distance score: computed by taking the median of the absolute errors. This objective is more robust against large outliers than the root mean squared error.
  • T-test score: Use the t-test feature extractor to select the features which are likely to be the most meaningful ones.

Updated solution score figure

We also updated the solution score figure. We now show the scores of the latest solutions tested by AutoML (light green), the current best score (orange) and the limit score (dark green), i.e. the maximum (or minimum) score which is theoretically possible.

Readme for Solutions

When you download a Solution from AutoML – a fully trained AI model – you want to have all the information about this particular Solution at your fingertips. That’s why we now provide an HTML README file with each Solution providing important information such as:

  • Which Dataset was used
  • Which Workflow generated the Solution
  • What AI model and Feature Extractors were used
  • What the Objective of the workflow was, and what the Solution’s score was
  • How to install all requirements
  • How to run the Online and Batch Clients

Other improvements and bug fixes

We made a range of fixes and improvements to AutoML. Some major improvements are:

  • Reduced the size of the files which AutoML generates while handling data.
  • Reduced the installation dependencies.
  • Support for mail-based password reset has been removed for security reasons.
  • Added helper text to the workflow creation screens.