Data-independent Random Projections from the feature-space of the homogeneous polynomial kernel

López-Sánchez, Daniel; Arrieta, Angélica González; Corchado, Juan M.

doi:10.1016/j.patcog.2018.05.003

Cited by 8 publications

(23 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, they discarded this idea because directly applying RP to the bilinear descriptors would involve storing a large projection matrix and explicitly computing the bilinear descriptors in the first place. However, recent advances in the intersection of kernel methods and Random Projection [20,[29][30][31] have made it possible to efficiently perform Random Projections from the feature spaces of different kernel functions in an efficient manner. In particular, an efficient method to approximate a Random Projection for polynomial kernels was introduced in [20] .…”

Section: Related Workmentioning

confidence: 99%

“…However, recent advances in the intersection of kernel methods and Random Projection [20,[29][30][31] have made it possible to efficiently perform Random Projections from the feature spaces of different kernel functions in an efficient manner. In particular, an efficient method to approximate a Random Projection for polynomial kernels was introduced in [20] . This paper adapts the ideas presented in [20] to make bilinear CNNs less computationally demanding by approximating a Random Projection of the bilinear descriptor.…”

Section: Related Workmentioning

confidence: 99%

“…In this paper, we further develop the idea of using Random Projection to reduce the dimension of the bilinear CNN descriptor. In particular, we propose adapting an existing kernelized variant of Random Projection [20] to efficiently project bilinear descriptors to a lower dimension without ever having to explicitly compute the high-dimensional bilinear descriptor itself. By implicitly computing a Random Projection of the bilinear descriptor, our method generates a low-dimensional feature vector that captures much of the discriminative information of the full bilinear descriptor, while resulting in models with a much lower number of parameters.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Compact bilinear pooling via kernelized random projection for fine-grained image categorization on low computational power devices

2020

Self Cite

View full text Add to dashboard Cite

Bilinear pooling is one of the most popular and effective methods for fine-grained image recognition. However, a major drawback of Bilinear pooling is the dimensionality of the resulting descriptors, which typically consist of several hundred thousand features. Even when generating the descriptor is tractable, its dimension makes any subsequent operations impractical and often results in huge computational and storage costs. We introduce a novel method to efficiently reduce the dimension of bilinear pooling descriptors by performing a Random Projection. Conveniently, this is achieved without ever computing the high-dimensional descriptor explicitly. Our experimental results show that our method outperforms existing compact bilinear pooling algorithms in most cases, while running faster on low computational power devices, where efficient extensions of bilinear pooling are most useful.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Compact bilinear pooling via kernelized random projection for fine-grained image categorization on low computational power devices

2020

Self Cite

View full text Add to dashboard Cite

show abstract

“…Existing kernels are either handcrafted (such as gaussian, histogram intersection, etc. [4,47,76,22,70]) or trained using multiple kernels [41,42,43,44,45] and explicit kernel maps [34,35,36,38,39,40] as well as their deep variants 1 [48,49,50,51,53,54,55,37,17].…”

Section: Introductionmentioning

confidence: 99%

“…In this proposed framework, the learned support vectors act as kernel parameters and make it possible to map input data to multiple kernel features prior to their classification (as also achieved in [34,35,36,38,39,40]). Nevertheless, the proposed framework is conceptually different from these related methods; the latter consider the support vectors fixed/taken from training data in order to design explicit kernel maps prior to learn parametric SVMs while in our method, the support vectors are optimized to better fit the task at hand.…”

Section: Introductionmentioning

confidence: 99%

Deep Total Variation Support Vector Networks

Sahbi

2019

2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)

View full text Add to dashboard Cite

Support vector machines (SVMs) have been successful in solving many computer vision tasks including image and video category recognition especially for small and mid-scale training problems. The principle of these nonparametric models is to learn hyperplanes that separate data belonging to different classes while maximizing their margins. However, SVMs constrain the learned hyperplanes to lie in the span of support vectors, fixed/taken from training data, and this reduces their representational power and may lead to limited generalization performances. In this paper, we relax this constraint and allow the support vectors to be learned (instead of being fixed/taken from training data) in order to better fit a given classification task. Our approach, referred to as deep total variation support vector machines, is parametric and relies on a novel deep architecture that learns not only the SVM and the kernel parameters but also the support vectors, resulting into highly effective classifiers. We also show (under a particular setting of the activation functions in this deep architecture) that a large class of kernels and their combinations can be learned. Experiments conducted on the challenging task of skeletonbased action recognition show the outperformance of our deep total variation SVMs w.r.t different baselines as well as the related work.

show abstract