Data

Tools

Indexes

Applications

Experiments

Open Source Science

siemens

Siemens Mobility’s AI Dependability Assessment Challenge

https://github.com/ericbill21/siemens

Science Score: 26.0%

This score indicates how likely this project is to be science-related based on various indicators:

  • ○
    CITATION.cff file
  • ✓
    codemeta.json file
    Found codemeta.json file
  • ✓
    .zenodo.json file
    Found .zenodo.json file
  • ○
    DOI references
  • ○
    Academic publication links
  • ○
    Academic email domains
  • ○
    Institutional organization owner
  • ○
    JOSS paper metadata
  • ○
    Scientific vocabulary similarity
    Low similarity (7.1%) to scientific vocabulary

Keywords

artificial-intelligence classification classifier dataset deep-learning machine-learning
Last synced: 6 months ago · JSON representation

Repository

Siemens Mobility’s AI Dependability Assessment Challenge

Basic Info
Statistics
  • Stars: 0
  • Watchers: 2
  • Forks: 0
  • Open Issues: 0
  • Releases: 0
Topics
artificial-intelligence classification classifier dataset deep-learning machine-learning
Created almost 5 years ago · Last pushed almost 4 years ago
Metadata Files
Readme Citation

readme.md

Siemens Mobilitys AI Dependability Assessment Challenge

Article about the challenge published by the RWTH Innovation Center

Dependability and reliability of machine learning classification systems are getting increasingly important as machine learning techniques are being applied in more and more domains, including safety-critical areas. The task of the Siemens AI Dependability Assessment Student's Challenge is to provide a classifier for a given classification tasks as well as to determine an as-accurate-as-possible misclassification probability estimate.

In this paper, we describe a machine learning model that has been optimized for the safest classification possible. A multilayer feed-forward neural network was trained on the three given datasets. We introduce a custom loss function to reduce the number of safety-critical misclassifications. Additionally, we detect predictions that are less certain than a determined threshold value and decide to resort to a safer classification in these instances.

Moreover, we describe the approach taken to determine a probability estimate for misclassifications: As training data quality is crucial to model performance and dependability, we develop several interpretations of training data and use these as indicators to estimate model accuracy.

For the three given datasets, our approach provides reliable (though not guaranteed) misclassification rates no larger than 15.23%, 8.63%, and 5.45%, respectively.

Dataset C

alt text

Weighted misclassification probability for dataset C

alt text

Owner

  • Name: Eric Bill
  • Login: ericbill21
  • Kind: user
  • Location: Aachen
  • Company: RWTH Aachen

Computer Science Student

GitHub Events

Total
Last Year