PaperlayerView exact revision on arXiv

Source-linked AI summary

Online Learning: A Comprehensive Survey

Steven C. H. Hoi, Doyen Sahoo, Jing Lu, Peilin Zhao

arXiv:1802.02871v2cs.LG

TL;DR

Batch learning can be inefficient and poorly scalable for evolving data because new training data may require retraining. This survey systematically reviews online learning, develops a taxonomy based on feedback and supervision, and mainly examines supervised and partial-feedback methods. It synthesizes the field's theoretical foundations, applications, and open issues while noting that comprehensive categorization remains difficult.

  • Problem

    Batch learning can suffer expensive retraining costs and poor scalability when new data arrives continuously and evolves rapidly.

  • Method

    The paper conducts a systematic review of online learning literature and categorizes methods according to feedback information and supervision.

  • Results

    The survey organizes online learning work into major feedback-based categories and reviews its theoretical foundations, methods, applications, and ongoing research trends.

  • Takeaways & Limitations

    Online learning provides an incremental framework for updating models as sequential data arrives, supporting scalable machine learning for real-world data analytics.

  • Takeaways & Limitations

    The proposed taxonomy may not fully cover all existing online learning work because precise categorization is challenging and the research evolves rapidly.

Abstract

from arXiv · show

Online learning represents an important family of machine learning algorithms, in which a learner attempts to resolve an online prediction (or any type of decision-making) task by learning a model/hypothesis from a sequence of data instances one at a time. The goal of online learning is to ensure that the online learner would make a sequence of accurate predictions (or correct decisions) given the knowledge of correct answers to previous prediction or learning tasks and possibly additional information. This is in contrast to many traditional batch learning or offline machine learning algorithms that are often designed to train a model in batch from a given collection of training data instances. This survey aims to provide a comprehensive survey of the online machine learning literatures through a systematic review of basic ideas and key principles and a proper categorization of different algorithms and techniques. Generally speaking, according to the learning type and the forms of feedback information, the existing online learning works can be classified into three major categories: (i) supervised online learning where full feedback information is always available, (ii) online learning with limited feedback, and (iii) unsupervised online learning where there is no feedback available. Due to space limitation, the survey will be mainly focused on the first category, but also briefly cover some basics of the other two categories. Finally, we also discuss some open issues and attempt to shed light on potential future research directions in this field.

1. Introduction

Online learning addresses the scalability challenges of batch learning by updating models incrementally as data arrives sequentially. This survey reviews the literature, organizes techniques by feedback and supervision, and focuses mainly on supervised and partial-feedback settings.

  • 1. Introduction: Batch learning requires expensive retraining for new data, limiting efficiency and scalability as data grows and changes rapidly.These constraints are especially relevant to continuous data streams and large-scale applications.
  • 1. Introduction: Online learning updates models incrementally and instantly when new sequential data arrives, supporting efficient and scalable learning for high-velocity data.The approach is designed for data arriving continuously rather than a complete dataset available in advance.
  • 1. Introduction: The survey synthesizes online learning literature across machine learning, data mining, statistics, optimization, applied mathematics, artificial intelligence, and data science.It aims to distill core methodological ideas and applications while acknowledging that rapidly evolving research makes complete coverage difficult.
  • 1.2 Tasks and Applications: Online learning applies to classification, regression, bandits, clustering, recommender systems, learning to rank, and reinforcement learning.Applications include spam filtering, financial time series, advertising, product recommendation, and streaming-data clustering.
  • 1.3 Taxonomy: Its taxonomy groups online learning techniques according to feedback information and supervision, with foundations in learning, optimization, and game theory.Figure 1 summarizes this organization, including supervised online learning and other feedback settings.

2. Problem Formulations and Related Theory

The section formalizes online prediction as sequential decision-making with feedback, loss measurement, model updates, and regret minimization. It then reviews statistical learning theory, online convex optimization, and game theory as foundations for online learning.

  • 2.1 Problem Settings: Online binary classification proceeds sequentially: the learner receives an instance, predicts, observes the true label, measures loss, and updates its model.The update is intended to improve predictive performance on future instances.
  • 2.1 Problem Settings: Sublinear regret means the learner’s average performance approaches that of the best fixed model in hindsight.The section defines regret relative to the hindsight-optimal model and states that R_T=o(T) implies R(T)/T→0.
  • 2.2 Statistical Learning Theory: Statistical learning theory models data through an unknown joint distribution P(x,y) and seeks a predictor minimizing expected loss, or true risk.Because the distribution is unknown, empirical risk is estimated from a finite i.i.d. sample.
  • 2.2 Statistical Learning Theory: Empirical Risk Minimization selects a hypothesis from F by minimizing empirical error; with linear classifiers and hinge loss, this becomes an objective over the model parameters.ERM is presented as a theoretical basis for many machine learning algorithms.
  • 2.2 Statistical Learning Theory: Excess risk decomposes into estimation error from finite data and approximation error from restricting the model class.More data generally reduces estimation error, while greater model capacity can reduce approximation error but may increase estimation error.
  • 2.3 Convex Optimization Theory: Online convex optimization uses a convex feasible set S and convex loss ℓ_t, choosing w_t∈S each round to minimize regret in hindsight.The framework includes loss functions, feasible decisions, and a comparator w* minimizing cumulative convex objective.
  • 2.3 Convex Optimization Theory: The survey distinguishes first-order and second-order OCO methods, with OGD using gradient information and projection, and ONS using curvature information.OGD achieves O(√T) regret for bounded-gradient convex costs, whereas ONS has O(log T) regret under exp-concave costs but higher O(n^2) complexity than OGD’s O(n).
  • 2.3 Convex Optimization Theory: ONS projects under a matrix-induced norm rather than the Euclidean norm used by OGD.The projection is defined through the matrix A_t.

3.1 Overview

This section introduces online supervised learning as a family of algorithms for supervised tasks and focuses on linear online learning, where the target hypothesis is a linear function.

  • 3.1 Overview: Online supervised learning algorithms provide fundamental approaches and principles for supervised online learning tasks.The survey treats these algorithms as a core family within online learning.
  • 3.1 Overview: Linear online learning seeks a linear hypothesis f:X→Y; in binary classification, the classifier predicts sign(w·x_t).The weight vector w defines the learned classifier in a typically d-dimensional input space.

3.2 First-order Online Learning

This section surveys first-order online learning algorithms, including Perceptron, Winnow, Passive-Aggressive, and OGD methods, together with their update rules and theoretical properties.

  • 3.2 First-order Online Learning: First-order online learning algorithms exploit first-order information during the learning process.The section frames these methods as a principal family of linear online learning algorithms.
  • 3.2.1 Perceptron: The Perceptron is the oldest online learning algorithm and updates a binary classifier when an incoming prediction is incorrect.Its procedure initializes w_1, predicts from w_t·x_t, receives y_t, and changes the model after a mistake.
  • 3.2.1 Perceptron: Under separability with margin γ and bounded instance norm R, the Perceptron makes at most a margin-dependent number of mistakes, with larger γ giving a tighter bound.The displayed bound is described using R and γ, where γ is the minimum signed margin.
  • 3.2.1 Perceptron: The normalized Perceptron modifies the update rule and improves the original Perceptron’s mistake bound.The passage introduces the modification without stating the complete displayed values.
  • 3.2.2 Winnow: Winnow uses multiplicative updates for learning monotone disjunctions over binary inputs, promoting or demoting feature weights after errors.Its mistake bound is αk(log_α θ+1)+n/θ under the stated parameter and target-function conditions.
  • 3.2.3 Passive-Aggressive Algorithms: Passive-Aggressive learning keeps the updated classifier close to the previous one while forcing correction when hinge loss is nonzero.The regularized variants introduce C to balance passiveness against an aggressiveness slack term.
  • 3.2.3 Passive-Aggressive Algorithms: Unlike Perceptron, Passive-Aggressive algorithms update whenever loss is nonzero, even when classification is correct, and have comparable theoretical mistake bounds but often better empirical performance.The comparison is reported directly from the cited discussion.
  • 3.2.4 Online Gradient Descent: OGD updates by subtracting a learning-rate-scaled loss gradient and projecting the result onto a convex feasible set.OGD and PA have similar update forms, but OGD commonly uses a predefined learning-rate scheme while PA chooses an optimal per-round rate subject to C.

3.3 Second-Order Online Learning

Second-order online learning uses curvature or feature-geometry information in addition to gradients, improving optimization or robustness at higher computational cost. The section reviews whitened-Perceptron-inspired SOP, confidence-weighted methods, AROW, and SCW.

  • Second-order algorithms exploit first- and second-order information to accelerate convergence, but typically require higher computational complexity.
  • Second-Order Perceptron: SOP approximates the data correlation matrix from previously seen instances, extending the whitened Perceptron to the online setting.
  • Confidence Weighted Learning: Confidence Weighted learning models the classifier as a Gaussian distribution and constrains new predictions to achieve a predefined correctness probability.
  • Adaptive Regularization of Weight Vectors: AROW modifies CW for non-separable data by balancing distribution stability, prediction loss, and growing confidence in the parameters.
  • Adaptive Regularization of Weight Vectors: AROW is described as robust to noisy data and readily generalizable to collaborative filtering and online feature selection.
  • Soft Confidence-Weighted Learning: SCW retains adaptive margins through a probability formulation and is presented as more efficient and effective than AROW.

3.4 Prediction with Expert Advice

Prediction with expert advice chooses among experts while minimizing regret relative to them. The survey covers regret objectives, Hannan consistency, and multiplicative-weight algorithms including Weighted Majority and Hedge.

  • The learner assigns a distribution over experts, incurs their weighted average loss, and seeks to minimize regret against each expert.
  • Hannan consistency requires vanishing per-round regret, so the forecaster approaches the performance of the best expert in the long run.
  • Weighted Majority: Weighted Majority makes binary predictions from expert advice and decreases the weight of an expert after it makes a mistake.
  • Hedge: Hedge generalizes Weighted Majority by normalizing expert weights into a distribution and updating them multiplicatively according to disclosed losses.
  • Other Expert-Advice Algorithms: EWAF differs from Hedge because its loss compares the forecaster’s prediction with the true label rather than using the distribution’s inner product with expert losses.
  • Parameter-Free Algorithms: Parameter-free algorithms address the difficulty of choosing a learning rate beforehand, with early work achieving regret matching prior optimally tuned bounds.

3.5 Online Learning with Regularization

Regularization-based online learning targets high-dimensional data by inducing sparse models and controlling update behavior. The survey reviews truncation, FOBOS, RDA, and adaptive regularization, while noting geometry-related limitations.

  • Sparse online learning uses regularization to reduce computational time and space costs for high-dimensional data with many features.
  • Truncated Gradient Descent: Naive truncation rounds small coefficients to zero, but may discard useful low-frequency features and thereby hurt performance.
  • Truncated Gradient Descent: TGD reduces truncation aggressiveness and achieves time and space complexity linear in the number of nonzero features, independent of dimensionality d.
  • FOBOS: FOBOS combines an online gradient step with a projection incorporating a sparsity-inducing regularizer, and its ℓ1-norm form is a special case of TGD.
  • RDA: RDA uses averaged gradients, a global proximal function, and a constant truncation threshold, producing significantly sparser solutions than FOBOS.
  • Adaptive Regularization: Both FOBOS and RDA may fail to fully exploit the geometry of the underlying data distribution through their auxiliary strongly convex functions.

3.6 Online Learning with Kernels

Kernel-based online learning extends sequential classification to nonlinear functions through kernel representations, but unrestricted support-vector growth creates scalability problems. Budget methods control this growth through maintenance strategies such as removal.

  • Kernel-based online learning represents nonlinear classifiers through kernels and measures classification confidence by |f(x)|.
  • Online nonlinear learners seek low cumulative regret relative to the best hypothesis in the kernel hypothesis space.
  • Kernel Representation: A kernel classifier is expressed as a weighted sum over support vectors, whose nonzero coefficients determine the model representation.
  • Kernel Methods: The kernel trick replaces inner products with kernel evaluations, allowing linear online algorithms such as Perceptron and OGD to obtain kernelized variants.
  • Scalability Challenge: Online kernel learning can achieve higher accuracy than linear models, but its support-vector count may grow without bound, increasing computation and storage costs.
  • Budget Maintenance: Budget online kernel methods cap the support-vector set at B and maintain that budget by adding vectors when needed and removing existing ones when exceeded.
  • Support-Vector Removal: Removal strategies include random deletion and discarding the oldest vector, but simple rules may provide unsatisfactory learning results.

3.7 Online to Batch Conversion

Online-to-batch conversion connects online learning’s sequential regret guarantees to batch-style generalization performance under stated loss assumptions. The survey presents general and strongly convex cases, including logarithmic regret for OGD and corresponding generalization bounds.

  • Online classification receives instances sequentially, predicts with the current model, observes labels and losses, and aims to minimize cumulative regret.
  • Online-to-batch conversion relates an online algorithm’s regret to the generalization performance of a model derived from its iterates.
  • The strongly convex analysis assumes convex, Lipschitz, and λ-strongly convex losses over the feasible domain, with sequential examples drawn i.i.d.
  • Under the stated assumptions, conditional expected loss equals the population risk, enabling a high-probability generalization theorem via Freedman’s inequality.
  • OGD achieves regret RegA(T) ≤ L2 2λ(1 + ln T), while the averaged iterate supports a high-probability generalization guarantee.
  • Prior work extends conversion results to general convex losses and sharper concentration through exponential-moment and Bernstein-based analyses.

4. Applied Online Learning for Supervised Learning

The survey reviews supervised online methods adapted to cost-sensitive classification, AUC maximization, and metric learning. These methods address task-specific objectives while managing non-convexity, class imbalance, storage, or projection costs.

  • 4.2 Cost-Sensitive Online Learning: Cost-sensitive online learning addresses the misleading nature of accuracy on imbalanced data in applications such as anomaly, fraud, and intrusion detection.
  • 4.2 Cost-Sensitive Online Learning: PAUM uses separate class margins for cost-sensitive updates but does not directly optimize a predefined cost-sensitive measure.
  • 4.2 Cost-Sensitive Online Learning: CSOGD directly optimizes weighted sensitivity-specificity and weighted misclassification-cost objectives through convex surrogate hinge losses.
  • 4.3 Online AUC Maximization: Online AUC maximization replaces a non-convex indicator objective with convex surrogates, but direct pairwise optimization would require storing all received instances.
  • 4.3 Online AUC Maximization: Reservoir sampling maintains class-specific buffers that provide unbiased historical sketches, allowing AUC loss approximation with bounded storage before OGD or PA updates.
  • 4.6 Online Metric Learning: OASIS learns a bilinear similarity matrix from streaming triplets and avoids requiring positive semidefiniteness, reducing projection-related computational cost.

4.7 Online Collaborative Filtering

Online collaborative filtering predicts ratings sequentially and updates models from observed errors, with memory-based and model-based approaches addressing similarity, sparsity, and scalability challenges. The section also situates online ranking methods in pointwise and pairwise formulations.

  • 4.7 Online Collaborative Filtering: Online collaborative filtering receives user-item indices sequentially, predicts a rating, observes the true rating, and updates from the prediction loss.
  • 4.7 Online Collaborative Filtering: Its learning objective is to minimize cumulative RMSE or MAE over the sequence of observed ratings.
  • 4.7 Online Collaborative Filtering: Memory-based collaborative filtering computes item or user similarities and predicts ratings as similarity-weighted averages of neighboring ratings.
  • 4.7 Online Collaborative Filtering: Memory-based methods face data sparsity and inefficient similarity updates, motivating model-based methods such as matrix factorization.
  • 4.7 Online Collaborative Filtering: Model-based collaborative filtering approximates ratings with user-item vector dot products and applies regularized online optimization, including OGD and later adaptive or second-order variants.
  • Online Learning to Rank: Online learning-to-rank methods observe instances sequentially and are categorized into pointwise and pairwise approaches.
  • Online Learning to Rank: Prank combines a Perceptron weight vector with ordered thresholds, updating them when predicted and true ranks disagree.
  • Online Learning to Rank: OAP-BPM generates multiple diverse weight-and-threshold solutions during each iteration and combines them for a final solution.

5. Bandit Online Learning

Bandit online learning studies sequential decisions under partial feedback, focusing on balancing exploration with exploitation. The survey reviews stochastic, adversarial, Bayesian, linear, and combinatorial bandit algorithms and their regret guarantees.

  • Bandit foundations: Bandit online learning gives only partial feedback while the learner repeatedly chooses actions and seeks to maximize cumulative payoff.The learner observes the payoff or loss of the selected action rather than complete information about all actions.
  • Problem settings: Stochastic bandits assume i.i.d. rewards from fixed unknown distributions, whereas adversarial bandits allow rewards to depend on prior actions.The distinction determines whether the environment is stochastic or adaptive to the learner’s history.
  • Stochastic bandit algorithms: ϵ-Greedy exploits the empirically best arm with probability 1−ϵ and explores randomly with probability ϵ.The constant-ϵ version incurs linearly growing regret, motivating exploration schedules that decrease over time.
  • Stochastic bandit algorithms: UCB combines exploration and exploitation by selecting actions using optimistic confidence bounds that become tighter as estimates improve.Its expected regret is bounded by O(√(KT ln T)) in the stated worst-case result.
  • Bayesian bandits: Thompson Sampling maintains Bayesian posteriors over arm means, samples candidate means, and selects the arm with the largest sample.The survey describes Beta priors and posterior updates for Bernoulli rewards, along with theoretical regret results.
  • Adversarial and structured bandits: Adversarial, linear, and combinatorial bandit methods extend partial-feedback learning to adaptive environments and structured action spaces with specialized regret bounds.The reviewed results include an adversarial bound of O(√(TK ln K)) for Exp3 and bounds for structured bandits.

6. Online Active Learning

Online active learning reduces labeling demands by selectively querying labels during sequential prediction. The survey reviews first- and second-order selective-sampling methods, expert-advice approaches, and related classifier families.

  • Motivation and settings: Online active learning addresses settings where obtaining every true label is expensive or unnecessary by deciding which incoming instances to query.Selective sampling and expert-advice formulations are identified as two major settings.
  • First-order selective sampling: The Selective-sampling Perceptron uses a randomized Bernoulli rule to decide whether to request each label, then updates with the Perceptron rule when queried.The query probability depends on δ and the prediction-related quantity |p̂_t|.
  • First-order selective sampling: Under ∥x_t∥≤R, the Selective-sampling Perceptron has an expected mistake bound for any comparator w and an associated expected query count.The survey presents the bound in terms of margin loss and the number of queried labels.
  • First-order selective sampling: The selective-sampling Perceptron recovers the familiar standard Perceptron bound in the linearly separable special case.The stated special case uses the optimal parameter δ=R^2/2 and yields the bound (∥w∥R)^2/γ^2.
  • Related approaches: Other approaches use RLS-based learners, expert disagreement, or Passive-Aggressive extensions, but RLS-based methods are described as less truly online because they resemble batch learning.The survey also notes extensions to multi-class, cost-sensitive, and expert-advice settings.
  • Second-order selective sampling: Second-order selective-sampling Perceptrons incorporate correlation-matrix information into prediction and retain randomized querying, with separate mistake and query guarantees.The second-order method applies a linear transformation involving the accumulated correlation matrix before prediction.

7. Online Semi-supervised Learning

Online semi-supervised learning incorporates unlabeled data into sequential learning, including manifold-regularization and active semi-supervised approaches. The survey emphasizes the computational cost of maintaining similarities to all prior instances and reviews sparse approximations.

  • Scope and approaches: Online semi-supervised learning combines sequential updating with unlabeled data and includes branches that adapt batch methods or use active querying.The survey identifies online manifold regularization and online active semi-supervised learning as major directions.
  • Manifold regularization: Manifold regularization minimizes labeled-data loss, supervised regularization, and prediction differences between similar unlabeled instances.The objective explicitly separates labeled loss, a classic regularizer, and the manifold term.
  • Computational limitation: A straightforward Online Gradient Descent solution requires O(T^2) time and substantial memory because each new instance is compared with all previous instances.The similarity computation must retain all instances, creating both time and space costs.
  • Sparse approximations: Sparse approximations retain either a recent buffer of τ instances or s cluster centers obtained with a random projection tree.The buffer may discard informative older instances, while clustering reduces the retained representation.
  • Sparse approximations: Incremental k-center quantization provides another approximation by grouping neighboring points into representative points for constructing an approximate similarity graph.The resulting graph is used to find a harmonic solution in semi-supervised learning.
  • Transductive learning: Transductive online learning assumes the complete unlabeled instance sequence is available before online prediction begins.It still learns from an arbitrary sequence of labeled examples while receiving the unlabeled inputs in advance.

8. Online Unsupervised Learning

Online unsupervised learning covers clustering, dimension reduction, density estimation, and anomaly detection over unlabeled data streams. The survey reviews their main methods, applications, and scalability challenges.

  • Online unsupervised learning learns models from unlabeled data streams without explicit feedback and is organized into four major groups.The groups are online clustering, online dimension reduction, online anomaly detection, and online density estimation.
  • Online Clustering: Online clustering includes partitioning, density-based, hierarchical, and other incremental approaches for organizing streaming instances.Partitioning methods include STREAM-style processing under a memory budget, while hierarchical methods incrementally insert points into cluster trees.
  • Online Clustering: Density-based clustering handles arbitrary cluster shapes, avoids requiring the number of clusters in advance, and is less vulnerable to outliers than many conventional methods.It identifies dense regions separated by sparse regions.
  • Online Dimension Reduction: Online dimension reduction maps high-dimensional streaming instances into lower-dimensional representations to improve efficiency, compression, visualization, and practical applicability.Reviewed approaches include online PCA, ICA, Kernel-PCA, and incremental manifold learning such as ISOMAP and LLE.
  • Online Density Estimation: Online density estimation constructs probability-density estimates from observed streams, with kernel methods facing a curse of kernelization.Estimating density at a point can require evaluating the kernel against all previously observed data points; merging approaches address this with a memory budget and kernel merging.
  • Online Anomaly Detection: Online anomaly detection identifies abnormal behavior in streams and includes distance-, density-, clustering-, and statistical-based approaches.Distance-based methods commonly use sliding windows to detect anomalies in the current window.

9. Related Areas and Other Terminologies

The survey distinguishes online learning from related paradigms by comparing their data-processing settings, objectives, and theoretical foundations. Incremental, sequential, stochastic, adaptive, interactive, reinforcement, and continual learning overlap with online learning but are not identical to it.

  • Related terminology is context-dependent, and several concepts may be used interchangeably despite differing meanings across settings.The survey explicitly notes that some distinctions are subjective and vary by context.
  • Incremental Learning: Incremental learning processes streams under limited space and computation, but it may operate online one example at a time or batch-wise with multiple examples.Thus, incremental learning can be implemented in either online or batch manners.
  • Sequential Learning: Sequential learning trains on ordered data and structured outputs, and it can be solved with either batch or online algorithms.Sequence classification is a simpler special case whose target class vector is reduced to a single variable.
  • Stochastic Learning: Stochastic learning is closely related to online learning but is typically motivated by accelerating existing batch methods through stochastic optimization.Its defining connection is methodological rather than an identity with online learning.
  • Adaptive Learning: Adaptive learning addresses dynamically changing environments, but many approaches rely on heuristic modifications of batch algorithms and lack formal performance guarantees.The survey distinguishes this from methods grounded primarily in online learning theory and algorithms.
  • Other Related Areas: Interactive learning places users or domain experts in the loop, while reinforcement learning seeks policies maximizing cumulative rewards and continual learning addresses sequential tasks without forgetting old ones.Online learning can support interactive systems, but these fields have distinct objectives and settings.

10. Conclusions

The survey categorizes online learning by feedback and emphasizes supervised online learning while also covering limited-feedback and unsupervised settings. It concludes that substantial challenges remain for changing, large-scale, diverse, noisy, deep, and continual data streams.

  • Concluding Remarks: Online learning methods are grouped by feedback into supervised online learning, limited-feedback online learning, and unsupervised online learning.Supervised learning receives full feedback, limited-feedback learning receives partial feedback, and unsupervised learning receives no additional feedback.
  • Concluding Remarks: The survey focuses mainly on supervised online learning because it has received more research attention and extends naturally from batch supervised learning.The paper also briefly covers the other two categories.
  • Open Issues: Non-stationary streams remain challenging because concept drift can change target concepts unpredictably, while existing approaches often assume restricted drift patterns.The survey identifies a lack of formal theoretical frameworks or principled solutions for general concept drift.
  • Open Issues: Large-scale online learning remains non-trivial for real-world data with extremely high volume and velocity, motivating further work on parallel online learning.The challenge persists despite online learning’s efficiency and scalability advantages over batch learning.
  • Open Issues: Online learning must better address data variety across multiple sources and structured, semi-structured, and unstructured formats.Most existing studies focus on single-source structured data represented in vector spaces.
  • Open Issues: Data veracity remains under-addressed because noise, missing values, and incomplete information can affect online learning efficacy.Many conventional studies implicitly assume perfect-quality data and feedback.
  • Open Issues: Online deep learning and online continual learning introduce further challenges involving accuracy, efficiency, and catastrophic forgetting.The survey presents both as emerging directions for learning from continuously arriving streams.
Loading 1802.02871v2…