Learning to Learn without Gradient Descent by Gradient Descent

By Yutian Chen, Matthew W. Hoffman, Sergio Gomez Colmenarejo, Misha Denil, Timothy P. Lillicrap, Matthew Botvinick, Nando de Freitas. 2016

329 citations recorded by the source corpus.

Subfields: Meta & Continual Learning

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

Unsupervised Learning of Neural Networks to Explain Neural Networks (2018, background): “Meta-learning: Our study is also related to meta-learning [5, 1, 20, 34].”
Learning to reinforcement learn (2016, background): “Recent work by Chen et al. (2016) is particularly close in spirit to the work we have presented here, and can be viewed as treating the case of “inﬁnite bandits” using a meta-learning strategy broadly analogous to the one we have pursued.”
LFPT5: A Unified Framework for Lifelong Few-shot Language Learning Based on Prompt Tuning of T5 (2021, background): “(2020) show that a large-scale pre-trained model (a frozen GPT-3) can act as a black-box meta-learner (Chen et al., 2017) and yield impressive few-shot performance via manually designed prompts constructed with task descriptions and some canonical examples.”
Meta-Learning: A Survey (2018, neutral): “Later work generalizes this approach by training an optimizer on synthetic functions, using gradient descent (Chen et al., 2016).”
Causal Discovery with Reinforcement Learning (2019, background): “Neural combinatorial optimizer and reinforcement learning Recent advances in sequence-to-sequence learning [26] have motivated the use of neural networks for optimization in various domains [30, 35, 3].”
Meta-Learning with Warped Gradient Descent (2019, background): “…meta-learner should be model-agnostic and learn either an explicit optimiser (Ravi & Larochelle, 2016; Andrychowicz et al., 2016; Li & Malik, 2016; Chen et al., 2017) or an initialisation for a standard optimiser (Finn et al., 2017; Nichol et al., 2018; Flennerhag et al., 2019), such that a…”
Learning to Teach with Dynamic Loss Functions (2018, unknown): “Our work, and the more general L2T, leverages automatic techniques to bypass human prior knowledge as much as possible, which is in line with the principles of learning to learn and meta learning [43, 50, 2, 57, 37, 29, 10, 14].”
Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges (2021, methodology): “For example, Chen et al. (2017) proposed using an LSTM to learn how to update the weights of an NN.”
Learning to Predict Gradients for Semi-Supervised Continual Learning (2022, background): “Gradient is an essential element for updating the model in deep learning [8], [9], [39], [43], [44], [45], [46].”
Learning Gradient Descent: Better Generalization and Longer Horizons (2017, methodology): “In another recent paper (Chen et al., 2016), an RNN is used to take current position and value as input and outputs the next position, and it works well for black-box optimization and simple RL tasks.”
Discovery of Useful Questions as Auxiliary Tasks (2019, background): “…et al., 2018), unsupervised learning (Gupta et al., 2018; Hsu et al., 2018), few-shot model adaptation (Nagabandi et al., 2018), and improving the optimizers (Andrychowicz et al., 2016; Li & Malik, 2017; Ravi & Larochelle, 2017; Wichrowska et al., 2017; Chen et al., 2016; Gupta et al., 2018).”
Meta-Tracker: Fast and Robust Online Adaptation for Visual Object Trackers (2018, methodology): “[2, 8, 52, 31] attempted to replace hand-crafted optimization algorithms with meta-learned deep networks.”
Learning to Optimize: A Primer and A Benchmark (2021, methodology): “Examples include convex `1-minimization [75], neural network training [13], black-box optimization [53] and combinatorial optimization [104].”
Improving Spatiotemporal Self-Supervision by Deep Reinforcement Learning (2018, methodology): “Meta-Learning for Deep Neural Networks: Recently, methods have proposed ways to improve upon the classical training of neural networks by, for example, automatizing the selection of hyper-parameters [2, 8, 42, 58, 16, 37].”