Machine Learning has become the essential tool for automating tasks that consist in predicting the output associated to a certain input. However many modern algorithms are mainly developed for the simple cases of classification and regression. Structured prediction is the field concerned with predicting outputs consisting of complex objects such as graphs, orientations or sequences. While these objects are often of practical interest, they do not have many of the mathematical properties that allow to design principled and computationally feasible algorithms with traditional techniques. In this thesis we investigate and develop algorithms for learning manifold-valued functions in the context of structured prediction. Differentiable manifolds are a mathematical abstraction used in many domains to describe sets with continuous constraints and non-Euclidean geometric properties. By taking a structured prediction approach we show how to define statistically consistent estimators for predicting elements of a manifold, in constrast to traditional structured predition algorithms that are restricted to output sets with finite cardinality. We introduce a wide range of applications that leverage manifolds structures. Above all, we study the case of the hyperbolic manifold, a space suited for representing hierarchical data. By representing supervised datasets within hyperbolic space we show how it is possible to invent new concepts in a previously known hierarchy and show promising results in hierarchical classification. We also study how modern structured approaches can help with practical robotics tasks, either improving performances in behavioural pipelines or showing more robust predictions for constrained tasks. Specifically, we show how structured prediction can be used to tackle inverse kinematics problems of redundant robots, accounting for the constraints of the robotic joints. We also consider the task of biological motion detection and show that by leveraging the sequence structure of video streams we significantly reduce the latency of the application. Our studies are complemented by empirical evaluations on both synthetic and real data.

Structured Machine Learning for Robotics

MARCONI, GIAN MARIA
2020-04-30

Abstract

Machine Learning has become the essential tool for automating tasks that consist in predicting the output associated to a certain input. However many modern algorithms are mainly developed for the simple cases of classification and regression. Structured prediction is the field concerned with predicting outputs consisting of complex objects such as graphs, orientations or sequences. While these objects are often of practical interest, they do not have many of the mathematical properties that allow to design principled and computationally feasible algorithms with traditional techniques. In this thesis we investigate and develop algorithms for learning manifold-valued functions in the context of structured prediction. Differentiable manifolds are a mathematical abstraction used in many domains to describe sets with continuous constraints and non-Euclidean geometric properties. By taking a structured prediction approach we show how to define statistically consistent estimators for predicting elements of a manifold, in constrast to traditional structured predition algorithms that are restricted to output sets with finite cardinality. We introduce a wide range of applications that leverage manifolds structures. Above all, we study the case of the hyperbolic manifold, a space suited for representing hierarchical data. By representing supervised datasets within hyperbolic space we show how it is possible to invent new concepts in a previously known hierarchy and show promising results in hierarchical classification. We also study how modern structured approaches can help with practical robotics tasks, either improving performances in behavioural pipelines or showing more robust predictions for constrained tasks. Specifically, we show how structured prediction can be used to tackle inverse kinematics problems of redundant robots, accounting for the constraints of the robotic joints. We also consider the task of biological motion detection and show that by leveraging the sequence structure of video streams we significantly reduce the latency of the application. Our studies are complemented by empirical evaluations on both synthetic and real data.
30-apr-2020
machine learning; robotics; manifolds; structured prediction; inverse kinematics; deep learning; recurrent neural networks; statistical learning;
File in questo prodotto:
File Dimensione Formato  
phdunige_4315027.pdf

accesso aperto

Descrizione: Documento principale contenente la tesi
Tipologia: Tesi di dottorato
Dimensione 6.97 MB
Formato Adobe PDF
6.97 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11567/1006091
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact