Deep learning models "in the wild": detecting and characterizing model drift
Tesi esterna in azienda
Riferimenti FABRIZIO LAMBERTI
Riferimenti esterni Lia Morra
Gruppi di ricerca GR-09 - GRAphics and INtelligent Systems - GRAINS
Tipo tesi TESI IN COLLAB. CON AZIENDA
Descrizione The performance of machine learning models may degrade over time due to a phenomenon generally denoted as model, concept or data drift. In the specific field of computer vision, deep convolutional neural networks are generally trained assuming, implicitly or explicitly, that the process generating the images and their associated labels is stationary and does not change over time, an assumption which may be violated in practice. It is therefore important to identify when a model is becoming “stale” and may need retraining; at the same, it is not feasible in practice to continuously collect labels to evaluate performance, hence the system should be able to self-monitor its output and detect when its distribution changes.
The main goal of this research activity is to study possible solutions to evaluate and detect model drift for convolutional neural networks, targeting two main tasks: classification and content-based image retrieval. A secondary objective is to design a monitoring dashboard, based on a suitable set of Key Performance Indicators. The research will be carried out in collaboration with Reale Mutua Assicurazioni and evaluated on benchmarks and real-life datasets. Strong analytical skills required. Basic knowledge of machine learning or data analytics is required.
Vedi anche http://grains.polito.it/work.php
Scadenza validita proposta 01/04/2021 PROPONI LA TUA CANDIDATURA