A ‘Self-Referential’ Weight Matrix

By J. Urgen Schmidhuber. 1993

How researchers see this paper

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

In Model Architectures

Citations in the corpus

HyperNetworks (2016, background): “Even before the work on HyperNEAT and DCT, Schmidhuber (1992; 1993) has suggested the concept of fast weights in which one network can produce context-dependent weight changes for a second network.”
Learning to Learn with Generative Models of Neural Network Checkpoints (2022, methodology): “Rather than parameterizing update rules, neural networks can be used to directly output or modify other neural networks’ parameters (Schmidhuber, 1992; 1993; Hinton & Plaut, 1987).”
Graph HyperNetworks for Neural Architecture Search (2018, background): “Schmidhuber (1992; 1993) proposed “fast-weights”, where one network produces weight changes for another.”
Learning Rapid-Temporal Adaptations (2017, background): “Mitchell et al. [24] used the reward gradients from neural networks to better generalize to robot con-trol tasks, whereas Schmidhuber [33] discussed the use of network weight matrices themselves for continuous adaptation of an agent in dynamic environments.”
Tagger: Deep Unsupervised Perceptual Grouping (2016, background): “Similar approaches even allow for metalearning [25], that is, learning a learning algorithm.”
PowerPlay: Training an Increasingly General Problem Solver by Continually Searching for the Simplest Still Unsolvable Problem (2011, neutral): “P may itself be the set of weight matrices of a separate RNN call ed RNN2, computing tasks for RNN1, and modifications of RNN1, using te chniques for network-modifying networks as described in [26, 28, 27].”
On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models (2015, background): “They may represent current neural activations or fast weights [229, 232, 231] or other dynamic variables that may change during information processing.”
Optimal Ordered Problem Solver (2002, neutral): “A gradient-based metalearning technique (Schmidhuber, 1993) for continuous program spaces of diﬀerentiable recurrent neural networks (RNNs) was also designed to favor better learning algorithms; compare the remarkable recent success of the related but technically improved RNN-based metalearner by…”
One Big Net For Everything (2018, methodology): “Many RNN-like models can be used to build general computers, e.g., RNNs controlling push-down automata [9, 41] or other types of differentiable memory [20] including differentiable fast weights [63, 65], as well as closely related RNN-based meta-learners [66, 25, 51].”
Differentiable plasticity: training plastic neural networks with backpropagation (2018, methodology): “Schmidhuber (1993a) has pointed out that such homogenous-plasticity networks can in principle learn to produce any desired trajectory.”
Reinforcement Learning with Self-Modifying Policies (1998, background): “Self-modifying recurrent neural nets [18, 41] are able to run their own weight change algorithms.”
Discovering Reinforcement Learning Algorithms (2020, background): “Other approaches include simple approximation [25], RNNs with Hebbian learning [21, 22], and gradient preconditioning [12].”
Rapid Adaptation with Conditionally Shifted Neurons (2017, background): “deep LSTM model that adapts in both space and time. 4 Related Work Mitchell et al. [24] used the reward gradients from neural networks to better generalize to robot control tasks, whereas Schmidhuber [33] discussed the use of network weight matrices themselves for continuous adaptation of an agent i”
A Generative Model for Sampling High-Performance and Diverse Weights for Neural Networks (2019, background): “Neural networks that generate their own weights have been discussed in (Schmidhuber, 1993; Chang & Lipson, 2018).”
Hyperparameter Ensembles for Robustness and Uncertainty Quantification (2020, background): “Learning this mapping is an instance of learning an hypernetwork [61, 62, 29] and falls within the scope of bilevel optimization problems [14].”
State Aware Imitation Learning (2017, background): “Finally, we would like to point out the similarities our work shares with meta learning techniques that learn the gradients (e.g. [12]) or determine the weight updates (e.g. [22], [8]) for a neural network.”
Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions (2019, background): “First considered by Schmidhuber (1993; 1992), hypernetworks are functions mapping to the weights of a neural net.”