Agnostic active learning

By Maria-Florina Balcan, Alina Beygelzimer, J. Langford. 2006

578 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Computational Theory and Mathematics

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computational Theory and Mathematics

Citations in the corpus

Incremental Learning from Scratch for Task-Oriented Dialogue Systems (2019, background): “This is similar to active learning (Balcan et al., 2009; Dasgupta et al., 2005).”
Co-teaching: Robust training of deep neural networks with extremely noisy labels (2018, background): “However, it is commonly known that boosting and active learning are sensitive to outliers and noise, and a few wrongly selected instances can deteriorate the learning performance of the whole model [12, 3].”
The Relationship Between Agnostic Selective Classification Active Learning and the Disagreement Coefficient (2017, methodology): “This querying strategy, which is often termed “disagreementbased,” has been used in a number of stream-based AL algorithms such as A2 (Agnostic Active), developed in [14], RobustCAL, studied by the authors of [15, 16] and [17], or the general agnostic AL algorithm of [18].”
Better Optimism By Bayes: Adaptive Planning with Rich Models (2014, methodology): “More generally, our task is reminiscent of the case of active classiﬁcation (Balcan, Beygelzimer, and Langford 2009).”
Two faces of active learning (2011, background): “There are now several learning algorithms which, on a variety of canonical examples, provably yield signiﬁcantly lower label complexity than supervised learning [9, 13, 10, 2, 3, 7, 15, 16, 12].”
Exponential Regret Bounds for Gaussian Process Bandits with Deterministic Observations (2012, background): “An indication of what might be possible can be found in (Balcan et al., 2009), where regions of the version space are eliminated once they can be excluded with sufficiently high probability.”
A General Agnostic Active Learning Algorithm (2007, methodology): “separability, uniform input distribution) [1–8], or is generally computationally prohibitive [2, 4, 9].”
Margin Based Active Learning (2007, methodology): “A more general version, with theoretical analysis, was considered in [2].”
Loss Minimization and Parameter Estimation with Heavy Tails (2013, methodology): “ses where the loss is classiﬁ- cation error, the rates can be improved to O˜(log(1/L⋆)) by using active learning, i.e., allowing the estimator to adaptively sample Y|X for values of X of its choosing [11, 3, 7]. We note that in our setting, the cases that are hardest for our (or any) estimator, are ”
Optimal Value of Information in Graphical Models (2009, background): “While sample complexity bounds have been derived for some active learning problems ( c.f., Dasgupta, 2005; Balcan, Beygelzimer, & Langford, 2006), we are not aware of any active learning algorithms that perform provably optimal (even for restricted classes of problem instances).”
Hierarchical sampling for active learning (2008, background): “Indeed, there are canonical examples in which active learning provably yields exponentially lower label complexity than supervised learning (Cohn et al., 1994; Freund et al., 1997; Dasgupta, 2005; Balcan et al., 2006; Balcan et al., 2007; Castro & Nowak, 2007; Hanneke, 2007; Dasgupta et al., 2007).”
Analysis of Perceptron-Based Active Learning (2009, methodology): “For half-spaces under the uniform input distribution, in the realizable setting, the algorithm of Balcan et al. (2006) can be implementated efficiently, as shown by Balcan et al.”
Maximizing Fun by Creating Data with Easily Reducible Subjective Complexity (2013, background): “Original active learning is limited to supervised classification tasks, for example (Balcan et al. 2009; Cohn 1994; Fedorov 1972; Hwang et al. 1991; MacKay 1992; Plutowski et al. 1994; Seung et al. 1992), asking which data points to evaluate next to maximize information gain, typically (but not nece”
Active Learning as Non-Convex Optimization (2009, background): “Previous work in active learning has shown that active learning can give provably better label complexity than passive learning under certain distribution assumptions (Balcan et al., 2006; Dasgupta et al., 2007, 2005).”
Cyberspace Security Using Adversarial Learning and Conformal Prediction (2015, background): “We note that RHT expands on agnostic active (A2) learning [44] [45] that maintains both a current version space and a region of uncertainty.”
From Theories to Queries (2011, background): “One way of thinking about the problem is agnostic active learning (Balcan et al., 2006; Dasgupta et al., 2008), a framework which relaxes the assumption that the oracle’s labels are trustworthy, yet still has positive theoretical results.”