search query: @keyword online learning / total: 4
reference: 3 / 4
Author: | Calandra, Roberto |
Title: | An exploration of deep belief networks toward adaptive learning |
Publication type: | Master's thesis |
Publication year: | 2011 |
Pages: | ix + 47 Language: eng |
Department/School: | Tietotekniikan laitos |
Main subject: | Informaatiotekniikka (T-115) |
Supervisor: | Simula, Olli |
Instructor: | Montesino Pouzols, Federico ; Raiko, Tapani |
OEVS: | Electronic archive copy is available via Aalto Thesis Database.
Instructions Reading digital theses in the closed network of the Aalto University Harald Herlin Learning CentreIn the closed network of Learning Centre you can read digital and digitized theses not available in the open network. The Learning Centre contact details and opening hours: https://learningcentre.aalto.fi/en/harald-herlin-learning-centre/ You can read theses on the Learning Centre customer computers, which are available on all floors.
Logging on to the customer computers
Opening a thesis
Reading the thesis
Printing the thesis
|
Location: | P1 Ark Aalto 8673 | Archive |
Keywords: | machine learning artificial neural networks adaptive learning online learning deep belief networks generative model evolving intelligent system belief regeneration |
Abstract (eng): | It is becoming more and more important to have Intelligent Systems that can adapt during the operating phase in order to cope with environments of which we do not have a sufficiently complete prior knowledge. In this respect, Adaptive Learning tries to overcome the limitations of traditional Machine Learning techniques. It assumes in fact that the data, provided as a stream, do not necessarily maintain the same probability distribution or the same amount of "classes". Recently a new method called Deep Belief Networks (DBNs) that improves the performances of classical Multi-Layer Perceptron, has gained popularity in the traditional Machine Learning community. DBNs has proved to be effective on a wide range of applications and in many cases outperformed even application specific methods. The purpose of this thesis is thus, presumably, a first attempt to explore properties of the DBNs that could be utilized in order to extend the advantages of traditional DBNs to the field of adaptive learning. Using the generative capability of DBNs we show that it is possible to transfer, through the generated samples, the beliefs acquired to another DBN. We call this process Belief Regeneration. Making use of the generative capability and a classifier on top, we are able to generate pairs of samples and labels. With this generated labelled dataset, we can extend the Belief Regeneration also to regenerate a classifier. This process, even if with loss of classification accuracy, proves to be possible and feasible. We then introduce a novel method for adaptive classifiers based on DBNs. Using the process of Belief Regeneration we are able to generate observations that together with the novel ones can be used to re-train our classifier. While this method is far from the state-of-the-art, in terms of classification accuracy, it is a proof-of-concept as it virtually eliminates the need of storing previous observations. |
ED: | 2011-10-28 |
INSSI record number: 42904
+ add basket
INSSI