Multi-Similarity Loss with General Pair Weighting for Deep Metric Learning

Wang, Xun; Han, Xintong; Wei-ling, Huang; Dong, Dengke; Scott, Matthew R.

doi:10.48550/arxiv.1904.06627

Cited by 11 publications

(2 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…We trained a Remora model using Multi-Similarity loss (α=2, β=50, base=0.5) and Multi-Similarity miner (ε=0.1) [64] as implemented in PyTorch-Metric-learning. Triplets provided by the miner were filtered to only contain samples from the same 5-mer in an effort to force the model to compare the two labels in the same sequence context.…”

Section: Methodsmentioning

confidence: 99%

Direct detection of 8-oxo-dG using nanopore sequencing

Pagès-Gallego,

van Soest,

Besselink

et al. 2024

Preprint

View full text Add to dashboard Cite

Genomic DNA is constantly subjected to oxidative damage, which is thought to be one of the major drivers of cancer and age-dependent decline. The most prominent consequence is the modification of guanine into 8-hydroxyguanine (8-oxo-dG), which has important mutagenic potential and plays a role in methylation-mediated gene regulation. Methods to simultaneously detect and quantify 8-oxo-dG within its genomic context have been lacking; mainly because these methods rely on indirect detection or are based on hydrolysis of the DNA. Nanopore sequencing has been deployed for the direct detection of base-modifications like cytosine methylation during sequencing. However, currently there is no model to detect 8-oxo-dG by nanopore sequencing due to the lack of training data. Here, we developed a strategy based on synthetic oligos to create long DNA molecules with context variability for effective deep learning and nanopore sequencing. Moreover, we showcase a training approach suitable to deal with the extreme scarceness of 8-oxo-dG compared to canonical G to enable specific 8-oxo-dG detection. Applied to an inducible tissue culture system for oxidative DNA damage, our approach reveals variable 8-oxo-dG distribution across the genome, a dissimilar context pattern to C>A mutations, and concurrent 5-mC depletion within a 2-kilobase window surrounding 8-oxo-dG sites. These findings not only underscore the potential of nanopore sequencing in epigenetic research, but also shed light on 8-oxo-dG's role in genomic regulation. By simultaneously measuring 5-mC and 8-oxo-dG at single molecule resolution, our study provides insights into the functional interplay between these DNA modifications. Moreover, our approach using synthetic oligos to generate a ground truth from machine learning modification calling could be applied to any other DNA modification. Overall, our work contributes to advancing the field of epigenetics and highlights nanopore sequencing as a powerful tool for studying DNA modifications.

show abstract

Section: Methodsmentioning

confidence: 99%

Direct detection of 8-oxo-dG using nanopore sequencing

Pagès-Gallego,

van Soest,

Besselink

et al. 2024

Preprint

View full text Add to dashboard Cite

show abstract

“…Jedna od pokušaja da se ovaj problem spore konvergencije reši jeste funkcija cilja više sličnosti (eng. Multisimilarity loss) (Wang, Han, Huang, Dong, Scot 2020 [3]) koji odabira teške uzorke u procesu obučavanja.…”

Section: Funkcija Ciljaunclassified

Primena Metričkog Učenja U Poljoprivredi

Tumbas¹

2023

Zbornik radova FTN

View full text Add to dashboard Cite

show abstract

Multi-level Distance Regularization for Deep Metric Learning

Kim¹,

Park

2021

AAAI

View full text Add to dashboard Cite

We propose a novel distance-based regularization method for deep metric learning called Multi-level Distance Regularization (MDR). MDR explicitly disturbs a learning procedure by regularizing pairwise distances between embedding vectors into multiple levels that represents a degree of similarity between a pair. In the training stage, the model is trained with both MDR and an existing loss function of deep metric learning, simultaneously; the two losses interfere with the objective of each other, and it makes the learning process difficult. Moreover, MDR prevents some examples from being ignored or overly influenced in the learning process. These allow the parameters of the embedding network to be settle on a local optima with better generalization. Without bells and whistles, MDR with simple Triplet loss achieves the-state-of-the-art performance in various benchmark datasets: CUB-200-2011, Cars-196, Stanford Online Products, and In-Shop Clothes Retrieval. We extensively perform ablation studies on its behaviors to show the effectiveness of MDR. By easily adopting our MDR, the previous approaches can be improved in performance and generalization ability.

show abstract

Multi-Similarity Loss with General Pair Weighting for Deep Metric Learning

Cited by 11 publications

References 1 publication

Direct detection of 8-oxo-dG using nanopore sequencing

Direct detection of 8-oxo-dG using nanopore sequencing

Primena Metričkog Učenja U Poljoprivredi

Multi-level Distance Regularization for Deep Metric Learning

Contact Info

Product

Resources

About