Learning with Hilbert–Schmidt independence criterion: A review and new perspectives

Wang, Tinghua; Dai, Xiaolu; Liu, Yuze

doi:10.1016/j.knosys.2021.107567

Cited by 18 publications

(11 citation statements)

References 137 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…2.4 and 2.5 to inner products from reproducing kernel Hilbert spaces. See [46,47] for more details. HSIC is formulated as follows:…”

Section: Similarity Between Representationsmentioning

confidence: 99%

See 1 more Smart Citation

Cultivating Diversity: A Comparison of Diversity Objectives in Neuroevolution

Reilstad,

Ellefsen

2024

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Inspired by biological evolution's ability to produce the complexity that is human brains, neuroevolution utilizes evolutionary algorithms for optimizing the hyperparameters and structure of neural networks. However, evolutionary algorithms fail to produce the same type of diversity as biological evolution can with the abundant range of adaptable and complex organisms in nature. Encouraging diversity in neuroevolution has seen increased interest in recent years with methods such as Novelty Search and Quality-Diversity optimization. Another promising, but less explored approach, is to explicitly encourage diversity with an additional diversity objective. There is, however, a lack of knowledge regarding the relationship between the type of diversity encouraging objective and the characteristics of the targeted problem. For instance, should a diversity of brain structures, behaviors, or neural firing patterns be encouraged when optimizing a walking robot? Discussion6.1 Differences between objectives of the same type of diversity . 6.1.1 Ad hoc behavioral diversity outperforms generic behavioral diversity .

show abstract

“…2.4 and 2.5 to inner products from reproducing kernel Hilbert spaces. See [46,47] for more details. HSIC is formulated as follows:…”

Section: Similarity Between Representationsmentioning

confidence: 99%

“…The original purpose of HSIC was to determine the statistical independence of two sets of variables but has since been used for various machine learning problems such as feature selection, clustering, dimensionality reduction, and kernel optimization [47].…”

Section: Similarity Between Representationsmentioning

confidence: 99%

Cultivating Diversity: A Comparison of Diversity Objectives in Neuroevolution

Reilstad,

Ellefsen

2024

Lecture Notes in Computer Science

View full text Add to dashboard Cite

show abstract

“…where K and L are kernel matrices derived from a set of input data, and HSIC is the Hilbert-Schmidt independence criterion (HSIC) that is used to compute a statistical dependence between two matrix kernels [45]. Apparently, CKA is a normalized version of HSIC that is invariant to uniform scaling.…”

Section: Experience Sharingmentioning

confidence: 99%

Hierarchical Deep Reinforcement Learning With Experience Sharing for Metaverse in Education

Hare

Tang

2023

IEEE Trans. Syst. Man Cybern, Syst.

View full text Add to dashboard Cite

Metaverse has gained increasing interest in education, with much of literature focusing on its great potential to enhance both individual and social aspects of learning. However, little work has been done to address the systems and technologies behind providing meaningful Metaverse learning. This article proposes a technical framework to address this research gap, where a hierarchical multiagent reinforcement learning approach with experience sharing is developed to augment the intelligence of nonplayer characters in Metaverse learning for personalization. The utility and benefits of the proposed framework and methodologies are demonstrated in Gridlock, a Metaverse learning game, as well as through extensive simulations.

show abstract

“…To approach the first question, we leverage the interplay of Hilbert-Schmidt independence criterion (HSIC) and orthogonal projection, hence the name HSIC-Bottleneck Orthogonalization (HBO). Taking a close look at both: HSIC is a non-parametric kernel-based technique utilized to assess the statistical (in)dependence of different layers, which has been widely adopted for various learning tasks (Wang, Dai, and Liu 2021) but is under-investigated in CL community (Wang et al 2023); And a basic idea behind the orthogonal projection is to regularize gradient update directions that do not disturb the weights of previous tasks (Zeng et al 2019). Based on them, the introduced HBO implements non-overwritten parameter updates facilitated by the HSIC-bottleneck training in an orthogonal space, where one can exploit readily available gradient updates by measuring nonlinear dependencies between the inputs and outputs.…”

Section: Introductionmentioning

confidence: 99%

Towards Continual Learning Desiderata via HSIC-Bottleneck Orthogonalization and Equiangular Embedding

Li,

Wang,

Chen

et al. 2024

AAAI

View full text Add to dashboard Cite

Deep neural networks are susceptible to catastrophic forgetting when trained on sequential tasks. Various continual learning (CL) methods often rely on exemplar buffers or/and network expansion for balancing model stability and plasticity, which, however, compromises their practical value due to privacy and memory concerns. Instead, this paper considers a strict yet realistic setting, where the training data from previous tasks is unavailable and the model size remains relatively constant during sequential training. To achieve such desiderata, we propose a conceptually simple yet effective method that attributes forgetting to layer-wise parameter overwriting and the resulting decision boundary distortion. This is achieved by the synergy between two key components: HSIC-Bottleneck Orthogonalization (HBO) implements non-overwritten parameter updates mediated by Hilbert-Schmidt independence criterion in an orthogonal space and EquiAngular Embedding (EAE) enhances decision boundary adaptation between old and new tasks with predefined basis vectors. Extensive experiments demonstrate that our method achieves competitive accuracy performance, even with absolute superiority of zero exemplar buffer and 1.02x the base model.

show abstract

Learning with Hilbert–Schmidt independence criterion: A review and new perspectives

Cited by 18 publications

References 137 publications

Cultivating Diversity: A Comparison of Diversity Objectives in Neuroevolution

Cultivating Diversity: A Comparison of Diversity Objectives in Neuroevolution

Hierarchical Deep Reinforcement Learning With Experience Sharing for Metaverse in Education

Towards Continual Learning Desiderata via HSIC-Bottleneck Orthogonalization and Equiangular Embedding

Contact Info

Product

Resources

About