Active Learning with Statistical Models

By David A. Cohn, Zoubin Ghahramani, Michael I. Jordan. 1996

For many types of learners one can compute the statistically "optimal" way to select data. We review how these techniques have been used with feedforward neural networks [MacKay, 1992; Cohn, 1994]. We then show how the same principles may be used to select data for two alternative, statistically-based learning architectures: mixtures of Gaussians and locally weighted regression. While the techniques for neural networks are expensive and approximate, the techniques for mixtures of Gaussians and locally weighted regression are both efficient and accurate.

Subfields: Statistical Learning Theory

How researchers see this paper

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

In Statistical Learning Theory

Citations in the corpus

Generative Adversarial Networks (2021, neutral): “The first two scenarios can be solved via supervised learning but they will cost a lot of human effort to enrich the labeled data or to use active learning approaches [11].”
Dataset Distillation (2018, methodology): “…is to summarize the entire dataset by a small subset, either by only using the “valuable” data for model training (Angelova et al., 2005; Felzenszwalb et al., 2010; Lapedriza et al., 2013) or by only labeling the “valuable” data via active learning (Cohn et al., 1996; Tong & Koller, 2001).”
Physics-Informed Neural Networks for Cardiac Activation Mapping (2020, background): “We can view the entropy as negative information and, in the case of a Gaussian posterior, it is proportional to the variance, which has also been proposed for active learning [22].”
Deep Reinforcement Learning for Autonomous Driving: A Survey (2020, background): “Curriculum learning is related to active learning [64] where the learning agent can focus on the interesting examples that stand near frontier of its knowledge and abilities.”
Learning Algorithms for Active Learning (2017, background): “Active learning is motivated by the observation that a model may perform better while training on less labeled data if it can choose the data on which it trains (Cohn et al., 1996).”
Active learning of inverse models with intrinsically motivated goal exploration in robots (2013, positive): “ the learning and exploration of robots in an autonomous manner. 1.2 Driving Autonomous Exploration Active learning algorithms can be considered as organized and constrained selfexploration processes [41, 30, 89, 105, 60]. In the regression setting, they are used to learn a regression mapping betwee”
The limits and potentials of deep learning for robotics (2018, background): “Active learning [21] can also comprise retrieving annotations from other sources such as the web.”
Deep Bayesian Active Learning with Image Data (2017, methodology): “Such frameworks for learning are referred to as active learning (Cohn et al., 1996) (also known as “experiment design” in the statistics literature), and have been used successfully in ﬁelds such as medical diagnosis, microbiology, and manufacturing (Tong, 2001).”
Bayesian SegNet: Model Uncertainty in Deep Convolutional Encoder-Decoder Architectures for Scene Understanding (2015, background): “Uncertainty is also immediately useful for other applications such as active learning [7], semi-supervised learning, or label propagation [1].”
A Wholistic View of Continual Learning with Deep Neural Networks: Forgotten Lessons and the Bridge to Active and Open World Learning (2020, methodology): “Clearly, such proof is beyond the current understanding of deep neural networks, but has been shown to be feasible in the context of parametric models such as Gaussian mixture models (Cohn et al., 1996) or naive Bayes (Roy and McCallum, 2001).”
Co-teaching: Robust training of deep neural networks with extremely noisy labels (2018, background): “This process is similar to boosting [11] and active learning [7].”
Machine learning of linear differential equations using Gaussian processes (2017, background): “Although not pursued here, this information is very useful in designing a data acquisition plan, often referred to as active learning [12,23,28], that can be used to optimally enhance our knowledge about the parametric linear equation under consideration.”
Interactive machine learning for health informatics: when do we need the human-in-the-loop? (2016, methodology): “While the techniques for neural networks are computationally expensive and approximate, the techniques for mixtures of Gaussians and locally weighted regression are both efﬁ-cient and accurate [74].”
Guiding Trajectory Optimization by Demonstrated Distributions (2017, methodology): “In this letter, we use locally weighted estimation to model such a conditional distribution [21], [22], however, our approach is not limited to specific regression methods and other regression methods such as Gaussian Mixture Regression and Gaussian Process Regression could be used instead.”
A comprehensive review of digital twin — part 1: modeling and twinning enabling technologies (2022, background): “4; (2) active learning where data are collected adaptively to refine data-driven models in a data-efficient manner (Cohn et al. 1996), with Bichon et al.”
Selection of Relevant Features and Examples in Machine Learning (1997, background): “Much of the empirical work on example selection (e.g., Gross, 1991; Cohn et al., 1996) has dealt with low-dimensional spaces, yet this approach clearly holds even greater potential for domains involving many irrelevant features.”
Human-in-the-loop machine learning: a state of the art (2022, background): “As it has been shown, CL is related to Transfer Learning and Multi-task Learning but is also connected to active learning (AL) (Cohn et al. 1996).”