• Publications
  • Influence
Understanding Negative Samples in Instance Discriminative Self-supervised Representation Learning
TLDR
An inconsistency in the existing analysis means theoretically, a large number of negative samples degrade supervised performance, while empirically, they improve the performance; this result is theoretically explained regarding negative samples. Expand
PAC-Bayesian Contrastive Unsupervised Representation Learning
TLDR
This work presents PAC-Bayesian generalisation bounds for CURL, which are then used to derive a new representation learning algorithm, and demonstrates that the algorithm achieves competitive accuracy, and yields non-vacuousgeneralisation bounds. Expand
PAC-Bayes Analysis of Sentence Representation
TLDR
It is shown that simple heuristics such as averaging and inverse document frequency weighted averaging are derived by the formulation of the PAC-Bayes bound, and proposed novel sentence vector learning algorithms are proposed on the basis of this analysis. Expand
Imputing Missing Values in EEG with Multivariate Autoregressive Models
TLDR
An EEG signal imputation method based on multivariate autoregressive (MAR) modeling and its iterative estimation and simulation, inspired by the multiple imputation procedure is proposed and evaluated with real data with artificial missing entries. Expand
Analyzing Centralities of Embedded Nodes
TLDR
This paper analyzes empirical distributions of several node centrality measures, such as PageRank, based on node classification results to give insights into the properties of embeddings, which can provide cues to improve embedding algorithms. Expand
Scalable Algorithm for Probabilistic Overlapping Community Detection
TLDR
This paper proposes a scalable overlapping community detection method by using the stochastic variational Bayesian training of latent Dirichlet allocation (LDA) models, which predicts sets of neighbor nodes with a community mixture distribution. Expand
Sharp Learning Bounds for Contrastive Unsupervised Representation Learning
TLDR
By regarding the contrastive loss as a downstream loss estimator, the theory not only improves the existing learning bounds substantially but also explains why downstream classification empirically improves with larger negative samples—because the estimation variance of the downstream loss decays with largernegative samples. Expand