End-to-End Incremental Learning

Castro, Francisco M.; Marín-Jiménez, Manuel J.; Guil, Nicolás; Schmid, Cordelia; Alahari, Karteek

doi:10.1007/978-3-030-01258-8_15

Cited by 635 publications

(474 citation statements)

References 27 publications

Supporting

Mentioning

470

Contrasting

Unclassified

Order By: Relevance

“…The goal is to build an efficient algorithm that takes as input (a) a model, M t that has been trained on the first t units of data {D 1 , · · · , D t } and (b) new data, D t+1 , to then output an updated model M t+1 . The efficiency of the algorithm can be achieved without direct access to the previous training data, {D 1 , · · · , D t }, while performance is maintained without catastrophically forgetting the previous model [16]. Ensemble learning algorithms can be used to achieve incremental learning.…”

Section: Related Work On Incremental Learningmentioning

confidence: 99%

“…Then, for each new batch of data, a SVM is trained on the new data and the support vectors from the previous learning step [18]. Deep learning approaches have been producing state-of-the-art results yet they continue to suffer from catastrophic forgetting, a dramatic decrease in overall performance when training with new data added incrementally [16]. [15] utilize Bayesian approach to update the deep learning model with new data.…”

Section: Related Work On Incremental Learningmentioning

confidence: 99%

“…[15] utilize Bayesian approach to update the deep learning model with new data. [16] uses new data and a small exemplar set from the old classes to update the model. Both methods show improvement on performance for incremental deep learning models.…”

Section: Related Work On Incremental Learningmentioning

confidence: 99%

“…Both methods show improvement on performance for incremental deep learning models. [15] works for models that have a determined number of classes which remains the same over time whereas [16] handles the cases that novel classes can be available when training. Typically, there are two types of incremental learning situations: 1) new training data D t+1 contains data from existing classes only; 2) new training data D t+1 has samples from both novel classes (new classes that were not known to the existing model) and existing classes [21].…”

Section: Related Work On Incremental Learningmentioning

confidence: 99%

See 3 more Smart Citations

Keeping up with the genomes: efficient learning of our increasing knowledge of the tree of life

Zhao

Cristian

Rosen

2019

Preprint

View full text Add to dashboard Cite

Current metagenomic taxonomic classifiers cannot computationally keep up with the pace of training data generated from genome sequencing projects, such as the exponentially-growing NCBI RefSeq bacterial genome database. When new reference sequences are added to training data, statically trained classifiers must be rerun on all data, resulting in a highly inefficient process. The rich literature of "incremental learning" addresses the need to update an existing classifier to accommodate new data without sacrificing much accuracy compared to retraining the classifier with all data. We demonstrate how classification improves over time by incrementally training a classifier on progressive RefSeq snapshots and testing it on: (a) all known current genomes (as a ground truth set) and (b) a real experimental metagenomic gut sample. We demonstrate that as a classifier model's knowledge of genomes grows, classification accuracy increases. The proof-of-concept naïve Bayes implementation, when updated yearly, now runs in 1/4 th of the non-incremental time with no accuracy loss. In conclusion, it is evident that classification improves by having the most current knowledge at its disposal. Therefore, it is of utmost importance to make classifiers computationally tractable to keep up with the data deluge. 1/19Figure 1. Number of updates in the NCBI Bacteria Genome database A: Accumulative number of genomes updates per year; B: compared with last year, the number of new updates per yearreads in metagenome sequencing data -uses aligners, read mappers, classifiers and other "base techniques" to solve this problem [6][7][8][9][10]. Taxonomic classification is usually one of the first steps in a metagenomic pipeline [11]. Once these organisms are identified, they are then used in downstream analyses, such as alpha/beta diversity measures, ordination, feature selection, phenotype classification, etc.However, while many methods have been proposed for taxonomic classification [12,13], the accuracy of these methods using different training databases has not been fully tested. This is an important issue, because as new genome data are generated, training data sets, such as the commonly used the NCBI Reference Sequence Database (RefSeq) will change over time. Nasko et al. [14] recently demonstrated that more reads are classified (as opposed to be assigned to an unclassified/unknown class) by the Kraken classifier with newer database versions. Nasko's analysis also suggested that changes in RefSeq over time may influence but not necessarily reduce the misclassification rate, with genus/species false positives shown to be 1% and 8% respectively. These metrics were, however, calculated only on a selection of 10 genomes. Accordingly, given that there is an ongoing rapid expansion of genomic data of microbial diversity, it is imperative to update taxonomic classifiers as new genomes/genes are discovered. Simply failing to update the model will result in lower accuracy due to incomplete knowledge. In addition, the way that most researchers tra...

show abstract

Section: Related Work On Incremental Learningmentioning

confidence: 99%

Section: Related Work On Incremental Learningmentioning

confidence: 99%

Section: Related Work On Incremental Learningmentioning

confidence: 99%

Section: Related Work On Incremental Learningmentioning

confidence: 99%

See 2 more Smart Citations

Keeping up with the genomes: efficient learning of our increasing knowledge of the tree of life

Zhao

Cristian

Rosen

2019

Preprint

View full text Add to dashboard Cite

show abstract

“…However, multitask models could help improve generalizability when introduced as an intermediary. Lifelong learning has been used to help models learn from separate but related tasks in stages or continuously and could be used to facilitate multitask training . Lifelong learning protocols have been used to transfer knowledge acquired on old tasks to new ones to improve generalization and facilitate model convergence.…”

Section: Introductionmentioning

confidence: 99%

A convolutional neural network algorithm for automatic segmentation of head and neck organs at risk using deep lifelong learning

et al. 2019

View full text Add to dashboard Cite

Purpose This study suggests a lifelong learning‐based convolutional neural network (LL‐CNN) algorithm as a superior alternative to single‐task learning approaches for automatic segmentation of head and neck (OARs) organs at risk. Methods and materials Lifelong learning‐based convolutional neural network was trained on twelve head and neck OARs simultaneously using a multitask learning framework. Once the weights of the shared network were established, the final multitask convolutional layer was replaced by a single‐task convolutional layer. The single‐task transfer learning network was trained on each OAR separately with early stoppage. The accuracy of LL‐CNN was assessed based on Dice score and root‐mean‐square error (RMSE) compared to manually delineated contours set as the gold standard. LL‐CNN was compared with 2D‐UNet, 3D‐UNet, a single‐task CNN (ST‐CNN), and a pure multitask CNN (MT‐CNN). Training, validation, and testing followed Kaggle competition rules, where 160 patients were used for training, 20 were used for internal validation, and 20 in a separate test set were used to report final prediction accuracies. Results On average contours generated with LL‐CNN had higher Dice coefficients and lower RMSE than 2D‐UNet, 3D‐Unet, ST‐ CNN, and MT‐CNN. LL‐CNN required ~72 hrs to train using a distributed learning framework on 2 Nvidia 1080Ti graphics processing units. LL‐CNN required 20 s to predict all 12 OARs, which was approximately as fast as the fastest alternative methods with the exception of MT‐CNN. Conclusions This study demonstrated that for head and neck organs at risk, LL‐CNN achieves a prediction accuracy superior to all alternative algorithms.

show abstract

Remembering for the right reasons: Explanations reduce catastrophic forgetting

et al. 2021

View full text Add to dashboard Cite

The goal of continual learning (CL) is to learn a sequence of tasks without suffering from the phenomenon of catastrophic forgetting. Previous work has shown that leveraging memory in the form of a replay buffer can reduce performance degradation on prior tasks. We hypothesize that forgetting can be further reduced when the model is encouraged to remember the evidence for previously made decisions. As a first step towards exploring this hypothesis, we propose a simple novel training paradigm, called Remembering for the Right Reasons (RRR), that additionally stores visual model explanations for each example in the buffer and ensures the model has "the right reasons" for its predictions by encouraging its explanations to remain consistent with those used to make decisions at training time. Without this constraint, there is a drift in explanations and increase in forgetting as conventional continual learning algorithms learn new tasks. We demonstrate how RRR can be easily added to any memory or regularization-based approach and results in reduced forgetting, and more importantly, improved model explanations. We have evaluated our approach in the standard and fewshot settings and observed a consistent improvement across various CL approaches using different architectures and techniques to generate model explanations and demonstrated our approach showing a promising connection between explainability and continual learning. Our code is available at https://github.com/ SaynaEbrahimi/Remembering-for-the-Right-Reasons.

show abstract

End-to-End Incremental Learning

Cited by 635 publications

References 27 publications

Keeping up with the genomes: efficient learning of our increasing knowledge of the tree of life

Keeping up with the genomes: efficient learning of our increasing knowledge of the tree of life

A convolutional neural network algorithm for automatic segmentation of head and neck organs at risk using deep lifelong learning

Remembering for the right reasons: Explanations reduce catastrophic forgetting

Contact Info

Product

Resources

About