Adaptive computation and machine learning

By Thomas G. Dietterich. 1998

How researchers see this paper

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

In Meta & Continual Learning

Citations in the corpus

Model-Based Reinforcement Learning for Atari (2019, background): “…et al., 2015; Finn et al., 2016; Finn & Levine, 2016; Ebert et al., 2017; Hafner et al., 2018; Piergiovanni et al., 2018; Rybkin et al., 2018; Sutton & Barto, 2017, Chapter 8), capturing important details of the scene (Ha & Schmidhuber, 2018), encouraging exploration (Oh et al., 2015),…”
A Survey on Neural Architecture Search (2019, methodology): “Temporal Difference Learning Approaches like SARSA (Rummery and Niranjan, 1994), TD-λ (Sutton and Barto, 1998), and Q-learning (Watkins, 1989) attempt to find this policy implicitly via approximating the optimal value functions.”
Learning Where to Attend with Deep Architectures for Image Tracking (2011, methodology): “Equation 2 is a Gaussian Process estimate of the reward surface and can be used to select a ﬁxation point for the next time step.”
A Kernel Loss for Solving the Bellman Equation (2019, background): “The instability becomes even harder to avoid, when training data (transitions) are sampled from an off-policy distribution, a situation known as the deadly triad (Sutton & Barto, 2018, Sec.”
Rethinking Parameter Counting in Deep Models: Effective Dimensionality Revisited (2020, methodology): “In this section, we provide further results the effective dimensionality, including its connection to both the bias-variance decomposition of predictive risk (Geman et al., 1992; Dobriban & Wager, 2018) as well as the Hilbert space norm of the induced kernel (Rasmussen & Williams, 2008).”
Causal discovery in heavy-tailed models (2019, background): “Peters et al. (2009) prove that an AR( p ) process is time reversible, that is, can be represented by an AR( p ) process in the reversed time direction, if and only if the noise is Gaussian.”
Prognostic Validation of a Neural Network Unified Physics Parameterization (2018, methodology): “The parameters 𝛼 are obtained by minimizing the loss function in (4) using stochastic gradient descent (Goodfellow et al., 2016; see Chapter 8).”