polito.it
Politecnico di Torino (logo)

Custom incremental learning approaches in real world scenarios

Enrico Postolov

Custom incremental learning approaches in real world scenarios.

Rel. Paolo Garza. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2021

[img]
Preview
PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (1MB) | Preview
Abstract:

Nowadays, machine learning approaches have become an integral part in a wide variety of technologies and use-cases, ranging through a lot of diverse fields: from healthcare to automotive, or from banking to supply chain optimization, arriving also at the point of directly influencing our everyday’s life (e.g. with smart-home devices, or even just with smartphones). In each of the aforementioned, usually, a big quantity of data needs to be collected, so that the various algorithms which are applied can provide better models and thus, give results that most of the times coincide with the "ground truth". To improve these performances, what is usually done by several companies developing machine learning-based solutions, is to cyclically train from zero their models with bigger quantities of data, which are accumulated by their platforms over time. As it can be imagined, the process of defining an entirely new classifier after every predefined time window, can be expensive in terms of time and, more importantly, resources. In an era where companies, especially start-ups, rely on external infrastructures provided "as a service" to deliver their products to the customers, saving is of vital importance, and thus the training process mentioned above can’t be the best way to go. In this work, we collaborated with a small company to try and solve this issue, exploiting those approaches defined as incremental: a single classifier is kept over time and updates due to the presence of new data are applied directly to that instance, giving it thus the possibility to adapt to possible changes in data without the need of starting from the beginning the learning mechanism. The classifiers taken into consideration are the following three: Naïve-Bayes classifier, Random Forest classifier and Extra-Trees classifier. From our results, the performances are comparable to the ones obtained following the standard approaches, outlining so a new path that can be pursued for the implementation of this type of solutions.

Relators: Paolo Garza
Academic year: 2020/21
Publication type: Electronic
Number of Pages: 116
Subjects:
Corso di laurea: Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering)
Classe di laurea: New organization > Master science > LM-32 - COMPUTER SYSTEMS ENGINEERING
Aziende collaboratrici: RestWorld s.r.l.
URI: http://webthesis.biblio.polito.it/id/eprint/18192
Modify record (reserved for operators) Modify record (reserved for operators)