Artificial Supervision in Machine Learning – Can AIs learn from AIs?

  • In one sentence

    In this thesis, we want to develop a machine learning (ML) model for computer vision and link it to several large-scale AIs (e.g., Google Cloud Vision AI) to research to what extent the custom model profits from the knowledge of large-scale AIs in specific situations.


    In recent years, the prevalence of machine learning models of most different scale increased significantly. Apart from infrequent ML models of outstanding complexity (e.g., Google Cloud Vision AI, IBM Maximo Visual Inspection, Microsoft Azure Computer Vision), developers around the world implement small- and mid-scale solutions for businesses and research. The reasonable question of why not permanently consulting very complex and accurate large-scale ML models can be answered with arguments such as limited financial resources in business and research, as well as concerns about data privacy. Therefore, up until today, developers and researchers train ML models using datasets and, at times, human knowledge (e.g., in active learning). But why not additionally benefit from knowledge of large-scale ML models for very selected data instances in order to improve the performance of a custom ML model for acceptable costs?

    Research Goal

    The objective of this thesis is to research the impact of incorporating additional knowledge from a large-scale ML model (e.g., Google Cloud Vision AI) on the prediction power and the model uncertainty of a custom machine learning model. Therefore, we propose a novel dimension in the supervision of ML models: Artificial supervision. In addition to supervision from data and human supervision (e.g., active learning), artificial supervision may be a third way to increase the prediction performance of custom ML models. To assess the performance of artificial supervision, we first build and train a ML model for computer vision using available image datasets. In the second stage, we access knowledge of several large-scale ML models in cases where the custom model faces, for instance, uncertainty during the prediction or low prediction performance. Here, we particularly want to study how to combine the knowledge of several large-scale AIs such that the custom model benefits most significantly (e.g., majority vote). Like this, we aim at improving the prediction of custom ML models. Finally, we want to evaluate whether the supervision of machine learning models with large-scale AIs can result in superior performance. In detail, you will...

     - apply your theoretical knowledge to a practical use case.
     - become an expert in deep learning.
     - develop, implement, and evaluate computer vision models. 


    We look forward to receiving your application because you...

     - have a solid technical understanding of machine and deep learning techniques. 
     - are proficient in Python programming (e.g., pandas, tensorflow or keras, scikit-learn). 
     - work in an independent and accurate way.



    Start: immediately | Duration: 6 Month


    We offer you a challenging research topic, close supervision, and the opportunity to develop practical and theoretical skills. If you are interested, please send your CV, transcript of records, and a brief letter of motivation to johannes.jakubik@kit.edu.