Bio-inspired Min-Nets Improve the Performance and Robustness of Deep Networks

Philipp Grüning; Erhardt Barth

Bio-inspired Min-Nets Improve the Performance and Robustness of Deep Networks

Abstract

Min-Nets are inspired by end-stopped cortical cells with units that output the minimum of two learned filters. We insert such Min-units into state-of-the-art deep networks, such as the popular ResNet and DenseNet, and show that the resulting Min-Nets perform better on the Cifar-10 benchmark. Moreover, we show that Min-Nets are more robust against JPEG compression artifacts. We argue that the minimum operation is the simplest way of implementing an AND operation on pairs of filters and that such AND operations introduce a bias that is appropriate given the statistics of natural images.

Originalsprache	Englisch
Titel	SVRHM 2021 Workshop NeurIPS
Erscheinungsdatum	2021
Publikationsstatus	Veröffentlicht - 2021

UN SDGs

Dieser Output leistet einen Beitrag zu folgendem(n) Ziel(en) für nachhaltige Entwicklung

SDG 3 – Gesundheit und Wohlergehen
SDG 4 – Qualitativ hochwertige Bildung
SDG 9 – Industrie, Innovation und Infrastruktur
SDG 11 – Nachhaltige Städte und Gemeinschaften
SDG 12 – Verantwortungsvoller Konsum und Produktion
SDG 14 – Lebensraum Wasser
SDG 15 – Lebensraum Land

Strategische Forschungsbereiche und Zentren

Zentren: Zentrum für Künstliche Intelligenz Lübeck (ZKIL)
Querschnittsbereich: Intelligente Systeme

DFG-Fachsystematik

4.43-05 Bild- und Sprachverarbeitung, Computergraphik und Visualisierung, Human Computer Interaction, Ubiquitous und Wearable Computing
machine learning
neural networks

Dokumente

https://webmail.inb.uni-luebeck.de/inb-publications/pdfs/GrBa21.pdf

Zitieren

@inproceedings{faccab347059498e9b8ea67f79b1c3a7,

title = "Bio-inspired Min-Nets Improve the Performance and Robustness of Deep Networks",

abstract = "Min-Nets are inspired by end-stopped cortical cells with units that output the minimum of two learned filters. We insert such Min-units into state-of-the-art deep networks, such as the popular ResNet and DenseNet, and show that the resulting Min-Nets perform better on the Cifar-10 benchmark. Moreover, we show that Min-Nets are more robust against JPEG compression artifacts. We argue that the minimum operation is the simplest way of implementing an AND operation on pairs of filters and that such AND operations introduce a bias that is appropriate given the statistics of natural images.",

author = "Philipp Gr{\"u}ning and Erhardt Barth",

year = "2021",

language = "English",

booktitle = "SVRHM 2021 Workshop NeurIPS",

}

TY - GEN

T1 - Bio-inspired Min-Nets Improve the Performance and Robustness of Deep Networks

AU - Grüning, Philipp

AU - Barth, Erhardt

PY - 2021

Y1 - 2021

N2 - Min-Nets are inspired by end-stopped cortical cells with units that output the minimum of two learned filters. We insert such Min-units into state-of-the-art deep networks, such as the popular ResNet and DenseNet, and show that the resulting Min-Nets perform better on the Cifar-10 benchmark. Moreover, we show that Min-Nets are more robust against JPEG compression artifacts. We argue that the minimum operation is the simplest way of implementing an AND operation on pairs of filters and that such AND operations introduce a bias that is appropriate given the statistics of natural images.

AB - Min-Nets are inspired by end-stopped cortical cells with units that output the minimum of two learned filters. We insert such Min-units into state-of-the-art deep networks, such as the popular ResNet and DenseNet, and show that the resulting Min-Nets perform better on the Cifar-10 benchmark. Moreover, we show that Min-Nets are more robust against JPEG compression artifacts. We argue that the minimum operation is the simplest way of implementing an AND operation on pairs of filters and that such AND operations introduce a bias that is appropriate given the statistics of natural images.

M3 - Conference contribution

BT - SVRHM 2021 Workshop NeurIPS

ER -

Bio-inspired Min-Nets Improve the Performance and Robustness of Deep Networks

Abstract

UN SDGs

Strategische Forschungsbereiche und Zentren

DFG-Fachsystematik

Dokumente

Fingerprint

Zitieren