A foundation for exact binarized morphological neural networks

Theodore Aouad; Hugues Talbot

Communication Dans Un Congrès Année : 2024

A foundation for exact binarized morphological neural networks

(1, 2) , (2, 1)

1
2

Theodore Aouad

Fonction : Auteur
PersonId : 1164145
IdHAL : theodore-aouad
ORCID : 0000-0002-6286-6884

Centre de vision numérique

OPtimisation Imagerie et Santé

Hugues Talbot

Fonction : Auteur
PersonId : 1939
IdHAL : hugues-talbot
ORCID : 0000-0002-2179-3498
IdRef : 181843056

OPtimisation Imagerie et Santé

Centre de vision numérique

Résumé

Training and running deep neural networks (NNs) often demands a lot of computation and energy-intensive specialized hardware (e.g. GPU, TPU...). One way to reduce the computation and power cost is to use binary weight NNs, but these are hard to train because the sign function has a non-smooth gradient. We present a model based on Mathematical Morphology (MM), which can binarize ConvNets without losing performance under certain conditions, but these conditions may not be easy to satisfy in real-world scenarios. To solve this, we propose two new approximation methods and develop a robust theoretical framework for ConvNets binarization using MM. We propose as well regularization losses to improve the optimization. We empirically show that our model can learn a complex morphological network, and explore its performance on a classification task. ∀I ∈ I(δ), χ(I) >

Mots clés

Deep Learning Neural Network Quantization Neural Network Binarization Mathematical Morphology

Domaines

Informatique [cs] Intelligence artificielle [cs.AI] Apprentissage [cs.LG] Mathématiques [math]

Fichier principal

lbqnn_2023_camera_ready1.pdf (690.09 Ko)

Levelset Separation.png (19.81 Ko)

Proj-positive-larger.png (29.15 Ko)

bcomplex.png (210 B)

bdiamond.png (210 B)

binarized-noisy-closest.png (15.59 Ko)

binarized-noisy-closest_old.png (273.04 Ko)

binarized-noisy-float.png (316.77 Ko)

binarized-noisy-learned.png (247.23 Ko)

bisel_schema.png (49.43 Ko)

bnn-1-bin.png (12.38 Ko)

bnn-1-float.png (57.45 Ko)

boxplot-mnist-binary.jpg (178.4 Ko)

boxplot-mnist-float.jpg (166.14 Ko)

bsquare.png (190 B)

conv_schema.png (45.24 Ko)

conv_schema.svg (34.69 Ko)

dcross.png (217 B)

disk.png (209 B)

forward-noisy.png (255.13 Ko)

grid-search-archi.png (8.25 Ko)

grid-search-bias.png (8.04 Ko)

grid-search-levelsets.png (7.39 Ko)

grid-search-weight.png (6.52 Ko)

hstick.png (183 B)

main.pdf (690.09 Ko)

main.synctex.gz (227.6 Ko)

noisti-bnn-1.png (34.83 Ko)

noisti-bnn-2.png (37.05 Ko)

noisti-bnn-model-1.png (17.81 Ko)

noisti-bnn-model-2.png (32.6 Ko)

noisy-input.png (29.55 Ko)

noisy-target.png (18.02 Ko)

scross.png (192 B)

time_projected.png (28.92 Ko)

training-curves.png (113.38 Ko)

Origine	Fichiers produits par l'(les) auteur(s)

Théodore Aouad : Connectez-vous pour contacter le contributeur

https://hal.science/hal-04231040

Soumis le : mardi 10 octobre 2023-11:36:07

Dernière modification le : jeudi 21 novembre 2024-14:14:02

Dates et versions

hal-04231040 , version 1 (10-10-2023)

Identifiants

HAL Id : hal-04231040 , version 1
ARXIV : 2401.03830

Citer

Theodore Aouad, Hugues Talbot. A foundation for exact binarized morphological neural networks. ICCV 2023 - International Conference on Computer Vision, workshop on Low Bit Quantized Neural Networks, Oct 2023, Paris, France. ⟨hal-04231040⟩

Exporter

BibTeX XML-TEI Dublin Core DC Terms EndNote DataCite

Collections

INRIA CVN CENTRALESUPELEC INRIA2 UNIV-PARIS-SACLAY GS-COMPUTER-SCIENCE HUB-IA

73 Consultations

74 Téléchargements

A foundation for exact binarized morphological neural networks

Résumé

Mots clés

Domaines

Dates et versions

Identifiants

Citer

Exporter

Collections

Altmetric

Partager