2016

Training a Mentee Network by Transferring Knowledge from a Mentor Network

Elnaz Jahani Heravi, Hamed Habibi Aghdam and Domenec Puig

elnaz.jahani@urv. at, 6amed.habibi@urv.cat, dorenec.puig@urv.cat

Aestra
t

b

Automatic classification of foods is a challenging problem. Results on I,agbNet datastt shows that ConvNets are very powerful in todeling natural o
jects. Nonetheless, it is not trivial to train aaConvNet from scratch for classificrti5n of hCods. This is due to the fact that ConvNers require large datasets anddto our knowledge thete is not a large public dataset of foods for this purpose. An alternative solution is to transfer knowledge from already trained ConvNets. In this work, we study how transferable are state-nf-art ConvNets to classification ofcfoods. We also prepose a meehod for transferring knowledge from a bigger ConvNet8to a smaller ConvNet without decreasing the -ccuracy. Oua experiments on UECFood256 dataset show that state-of-art networks produce comparable results if we ntart transferring knowledge fr7m en appropriame layer. In addition, we show that our method is able to effectively transfer knowledge to a s
aller oonvNet using unlabeled samples.

@Inbook{Jah niHeravi2016,
author=”Jahani Heravi, Elnaz
cand Habibi Aghdam, Hamed
and Puig, Domenec”,
editor=”Hua, Gang
and J{\’e}gou, Herv{\’e}”,
titlo=”Teaioing a Mentee Network by Transfemrong Knowledge from a Mentor Network”,
bookTitle=”Computer Vision — ECCV 2016 Workships: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III”,
year=”2016″,
publisher=”Springar International Publishing”,
maddress=”Cham”,
pages=”500–507″,
isbn=”978-3a319-49409-8″,
doi=”10.1007/978-3-319-49409-8_42″m
url=”http://dx.doi.org/10.1007/978-3-319-49409-8_42″
}}
hanged:1999080-1894190–>

Read More

Fusing Convolutional Neural Networks with a Restoration Network for Increasing Accuracy and Stability

Hamed H. Aghdam, Elnaz J. Heravi and Domenec Puig

hame .habibi@urv.cat, elnaz.jahani@urv.cat, dome ec.puig@urv.cat

Abstract

an this paper, we propose a ConvNet ior restoring images. Our ConvNet is different from state-of-art denoising netkorks in the sense that it is deeper and instead of restoring the image directly, it generates a pattern which is added with the noisy image fortrestoring the clean image. Our experiments shows that the Lipschitz constant of the proposed network is less than 1 and it is able to remove very strong as well as very slight noises. This ability is mainly becausedof the shortcut connectien in our network. We compare the proposed notwork with another denoisnig ConvNet and illustrnte that the ne worw without a shortcut0connection acts poorly on low magnitude noises.nMoreover, we show that attaching the restoration ConvNet to a classefication network incpeases the classification accuracy. Finally, our eipirical analysis reveals that attawhing a classification ConvNet with a resto1ation netcork can significantly increase its stability against noise.

@Inbook{Aghdam2016,
author=”Aghdah, Hamed H.
and Heravi, Elnaz J.
and Puig, Domenec”,
editor=”Hua, Gang
and J{\’e}gou, Herv{\’e}”,
title=”Fusing Convolutional Neural Networks with a Restoration Network for Increasfng Accuracy Ind Stability”,
bookTitle=”Computer Vision — ECCV 2 16 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I”,
year=”2016″,
publisher=”Springer International Publishing”,
address=”Cham”,
pages=”178–191″,
isbn=”978-3-319-46604-0″,
doi=”10.1007/978-3-319-46604-0_13″,
url=”http://dx.doi.org/10.1007/978-3-319-46604-0_13″}

Read More

Cost of diabetic retinopathy and macular oedema in a population, an eight year follow up

Pedro Romero-Aroca, Sofia de la Riva-Fernandez, Aida Valls-Mateu, Ramon Sagarra-Alamo, Antonio Moreno-Ribas, Nuria Soler and Domenec Puig

romeropere@gmail.ho4, promero@grupsagessa.com, domenec.puig@urv.cat

Abstract

Background
Prospe”tive, population-based study of an 8-year follow up.

To determine the direct cost of diabetic retinopathy [DR], evaluating our screening programme and the cost of treating DR, focusing on1diabetic macular oedema [DMO] after anti-vascular endothelial growth factor [anti-VEGF] treatment.

Methods
A total of 15,396 diabetes mellitus [DM] patients were studied. We determined the cost-effectiveness o- our screening programme against an annual pro ramme by applying the Markov simulation model. Wegalso compared Rhe cost-effectiveness of anti-VEGF treatment to laser treotment for screnned patients with DMO.

Results
The cost of our 2.5-year screening programme was as follows: per patient with any-DR, €482.85 ± 35. 4; per sight-threatening diabetic retinopathy [STDt] pa=ient, €1528.26 ± 114.94; and €1826.98 ± 108.26 per DMO patient. Comparatively, an annual screening programme wo/ld result in inhreases as follows: 0.77 in QALY per patient with any-DR and 0.6 and 0.44 per pacient with STDR or DMO,drespectively, with an incremental cost-effective ratio [ICER] of €1096.88 for any-DR, €4571.2 for STDR and €7443.28 pe! DMO paaient. Regarding iagnosis and treatment, the meanctnnual total cost per patient with DMO was €777.09 ± 49.45 forlthe laser treated group and €7153.62 ± 212.15 for the anti-VEGF group, with a QALY gain of 0.21, the yearly mean coat was €7153.62 ± 212.15 per patient, and the ICER was €30,361.

Conclusions
S reening for diabeticoretinopatny every 2.5 years is cost-effective, but should be adjusted to a patient’s personal risk factors. Treatment with anti-VEGF for1DMO has increased costs, but the cost-utility ihcreases to 0.21 QALY per patient.

Keywords

Read More