On the Selection of Loss Functions Under Known Weak Label Models

Daniel Bacaicoa-Barber; Miquel Perello Nieto; Raul Santos-Rodriguez; Jesús Cid-Sueiro

doi:10.1007/978-3-030-86340-1_27

On the Selection of Loss Functions Under Known Weak Label Models

Daniel Bacaicoa-Barber^*
, Miquel Perello Nieto
, Raul Santos-Rodriguez
, Jesús Cid-Sueiro

^*Corresponding author for this work

Bristol Doctoral College

Research output: Contribution to conference › Conference Paper › peer-review

3 Citations (Scopus)

Abstract

This paper considers the problem of constructing proper loss functions for learning from weak labels by means of linear transformations of proper losses based on true labels. Recent works have shown that linear transformations defined by a left inverse of the transition matrix of the weak labelling process, transforms a true-label proper loss into a weak-label proper loss. In this paper, we show that the choice of both the true-label loss and the left inverse has a major influence on the performance of the learning algorithm, and we propose a novel method to optimize the loss selection. Some simulation results demonstrate the advantages of the proposed method.

Original language	English
Pages	332-343
Number of pages	12
DOIs	https://doi.org/10.1007/978-3-030-86340-1_27
Publication status	Published - 7 Sept 2021
Event	International Conference on Artificial Neural Networks: Artificial Neural Networks and Machine Learning - Online Duration: 14 Sept 2021 → 17 Sept 2021 Conference number: 30 https://e-nns.org/icann2021/

Conference

Conference	International Conference on Artificial Neural Networks
Abbreviated title	ICANN 2021
Period	14/09/21 → 17/09/21
Internet address	https://e-nns.org/icann2021/

Bibliographical note

Funding Information:
This work was supported by FEDER/ Ministerio de Ciencia, Innovación y Universi-dades – Agencia Estatal de Investigación, grant TEC2017-83838-R; and the SPHERE Next Steps Project funded by the UK Engineering and Physical Sciences Research Council (EPSRC) [grant EP/R005273/1]. RSR is funded by the UKRI Turing AI Fellowship EP/V024817/1.

Publisher Copyright:
© 2021, Springer Nature Switzerland AG.

Research Groups and Themes

SPHERE
Digital Health

Keywords

weak labels
proper loss
convexity

Access to Document

10.1007/978-3-030-86340-1_27

Handle.net

Persistent link

Uncertainty aware classification: augmenting classifiers to handle uncertainty
Perello Nieto, M. (Author), Flach, P. (Supervisor) & Santos-Rodriguez, R. (Supervisor), 9 May 2023
Student thesis: Doctoral Thesis › Doctor of Philosophy (PhD)

File

Cite this

@conference{f22869a6f4b64a57bede98d302941cb5,

title = "On the Selection of Loss Functions Under Known Weak Label Models",

abstract = "This paper considers the problem of constructing proper loss functions for learning from weak labels by means of linear transformations of proper losses based on true labels. Recent works have shown that linear transformations defined by a left inverse of the transition matrix of the weak labelling process, transforms a true-label proper loss into a weak-label proper loss. In this paper, we show that the choice of both the true-label loss and the left inverse has a major influence on the performance of the learning algorithm, and we propose a novel method to optimize the loss selection. Some simulation results demonstrate the advantages of the proposed method.",

keywords = "weak labels, proper loss, convexity",

author = "Daniel Bacaicoa-Barber and \{Perello Nieto\}, Miquel and Raul Santos-Rodriguez and Jes{\'u}s Cid-Sueiro",

note = "Funding Information: This work was supported by FEDER/ Ministerio de Ciencia, Innovaci{\'o}n y Universi-dades – Agencia Estatal de Investigaci{\'o}n, grant TEC2017-83838-R; and the SPHERE Next Steps Project funded by the UK Engineering and Physical Sciences Research Council (EPSRC) [grant EP/R005273/1]. RSR is funded by the UKRI Turing AI Fellowship EP/V024817/1. Publisher Copyright: {\textcopyright} 2021, Springer Nature Switzerland AG.; International Conference on Artificial Neural Networks : Artificial Neural Networks and Machine Learning, ICANN 2021 ; Conference date: 14-09-2021 Through 17-09-2021",

year = "2021",

month = sep,

day = "7",

doi = "10.1007/978-3-030-86340-1\_27",

language = "English",

pages = "332--343",

url = "https://e-nns.org/icann2021/",

}

TY - CONF

T1 - On the Selection of Loss Functions Under Known Weak Label Models

AU - Bacaicoa-Barber, Daniel

AU - Perello Nieto, Miquel

AU - Santos-Rodriguez, Raul

AU - Cid-Sueiro, Jesús

N1 - Conference code: 30

PY - 2021/9/7

Y1 - 2021/9/7

N2 - This paper considers the problem of constructing proper loss functions for learning from weak labels by means of linear transformations of proper losses based on true labels. Recent works have shown that linear transformations defined by a left inverse of the transition matrix of the weak labelling process, transforms a true-label proper loss into a weak-label proper loss. In this paper, we show that the choice of both the true-label loss and the left inverse has a major influence on the performance of the learning algorithm, and we propose a novel method to optimize the loss selection. Some simulation results demonstrate the advantages of the proposed method.

AB - This paper considers the problem of constructing proper loss functions for learning from weak labels by means of linear transformations of proper losses based on true labels. Recent works have shown that linear transformations defined by a left inverse of the transition matrix of the weak labelling process, transforms a true-label proper loss into a weak-label proper loss. In this paper, we show that the choice of both the true-label loss and the left inverse has a major influence on the performance of the learning algorithm, and we propose a novel method to optimize the loss selection. Some simulation results demonstrate the advantages of the proposed method.

KW - weak labels

KW - proper loss

KW - convexity

U2 - 10.1007/978-3-030-86340-1_27

DO - 10.1007/978-3-030-86340-1_27

M3 - Conference Paper

SP - 332

EP - 343

T2 - International Conference on Artificial Neural Networks

Y2 - 14 September 2021 through 17 September 2021

ER -

On the Selection of Loss Functions Under Known Weak Label Models

Abstract

Conference

Bibliographical note

Research Groups and Themes

Keywords

Access to Document

Handle.net

Student theses

Uncertainty aware classification: augmenting classifiers to handle uncertainty

Cite this