Adapting by copying. Towards a sustainable machine learning
Despite the rapid growth of machine learning in the past decades, deploying automated decision making systems in practice remains a challenge for most companies. On an average day, data scientists face substantial barriers to serving models into production. Production environments are complex ecosys...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2021 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/671692 |
| Acceso en línea: | http://hdl.handle.net/10803/671692 |
| Access Level: | acceso abierto |
| Palabra clave: | Aprenentatge electrònic Aprendizaje electrónico Web-based instruction Algorismes Algoritmos Algorithms Presa de decisions Toma de decisiones Decision making Sistemes autoorganitzatius Sistemas autoorganizativos Self-organizing systems Ciències Experimentals i Matemàtiques 004 |
| Sumario: | Despite the rapid growth of machine learning in the past decades, deploying automated decision making systems in practice remains a challenge for most companies. On an average day, data scientists face substantial barriers to serving models into production. Production environments are complex ecosystems, still largely based on on-premise technology, where modifications are timely and costly. Given the rapid pace with which the machine learning environment changes these days, companies struggle to stay up-to-date with the latest software releases, the changes in regulation and the newest market trends. As a result, machine learning often fails to deliver according to expectations. And more worryingly, this can result in unwanted risks for users, for the company itself and even for the society as a whole, insofar the negative impact of these risks is perpetuated in time. In this context, adaptation is an instrument that is both necessary and crucial for ensuring a sustainable deployment of industrial machine learning. This dissertation is devoted to developing theoretical and practical tools to enable adaptation of machine learning models in company production environments. More precisely, we focus on devising mechanisms to exploit the knowledge acquired by models to train future generations that are better fit to meet the stringent demands of a changing ecosystem. We introduce copying as a mechanism to replicate the decision behaviour of a model using another that presents differential characteristics, in cases where access to both the models and their training data are restricted. We discuss the theoretical implications of this methodology and show how it can be performed and evaluated in practice. Under the conceptual framework of actionable accountability we also explore how copying can be used to ensure risk mitigation in circumstances where deployment of a machine learning solution results in a negative impact to individuals or organizations. |
|---|