A5102898754- Reinforcement Learning in Robotics
- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- Neural dynamics and brain function
- Artificial Intelligence in Games
- Advanced Graph Neural Networks
- Neural Networks and Applications
- Explainable Artificial Intelligence (XAI)
- Topic Modeling
- Adversarial Robustness in Machine Learning
- Advanced Bandit Algorithms Research
- Machine Learning and Data Classification
- Machine Learning and Algorithms
- Anomaly Detection Techniques and Applications
- Evolutionary Algorithms and Applications
- Bayesian Methods and Mixture Models
- Human Pose and Action Recognition
- Gaussian Processes and Bayesian Inference
- Opinion Dynamics and Social Influence
- Machine Learning in Materials Science
- Neural and Behavioral Psychology Studies
- Advanced Image and Video Retrieval Techniques
- Generative Adversarial Networks and Image Synthesis
- Memory and Neural Mechanisms
- Neural Networks and Reservoir Computing
DeepMind (United Kingdom)
2013-2022
Google (United States)
2013-2021
Google (United Kingdom)
2016-2019
University College London
2010-2015
Oxford Centre for Computational Neuroscience
2011
Deep neural networks (NNs) are powerful black box predictors that have recently achieved impressive performance on a wide spectrum of tasks. Quantifying predictive uncertainty in NNs is challenging and yet unsolved problem. Bayesian NNs, which learn distribution over weights, currently the state-of-the-art for estimating uncertainty; however these require significant modifications to training procedure computationally expensive compared standard (non-Bayesian) NNs. We propose an alternative...
We introduce a new, efficient, principled and backpropagation-compatible algorithm for learning probability distribution on the weights of neural network, called Bayes by Backprop. It regularises minimising compression cost, known as variational free energy or expected lower bound marginal likelihood. show that this kind regularisation yields comparable performance to dropout MNIST classification. then demonstrate how learnt uncertainty in can be used improve generalisation non-linear...
For artificial general intelligence (AGI) it would be efficient if multiple users trained the same giant neural network, permitting parameter reuse, without catastrophic forgetting. PathNet is a first step in this direction. It network algorithm that uses agents embedded whose task to discover which parts of re-use for new tasks. Agents are pathways (views) through determine subset parameters used and updated by forwards backwards passes backpropogation algorithm. During learning, tournament...
Efficient exploration in complex environments remains a major challenge for reinforcement learning. We propose bootstrapped DQN, simple algorithm that explores computationally and statistically efficient manner through use of randomized value functions. Unlike dithering strategies such as epsilon-greedy exploration, DQN carries out temporally-extended (or deep) exploration; this can lead to exponentially faster demonstrate these benefits stochastic MDPs the large-scale Arcade Learning...
We introduce NoisyNet, a deep reinforcement learning agent with parametric noise added to its weights, and show that the induced stochasticity of agent's policy can be used aid efficient exploration. The parameters are learned gradient descent along remaining network weights. NoisyNet is straightforward implement adds little computational overhead. find replacing conventional exploration heuristics for A3C, DQN dueling agents (entropy reward $ε$-greedy respectively) yields substantially...
The practice of mathematics involves discovering patterns and using these to formulate prove conjectures, resulting in theorems. Since the 1960s, mathematicians have used computers assist discovery formulation conjectures1, most famously Birch Swinnerton-Dyer conjecture2, a Millennium Prize Problem3. Here we provide examples new fundamental results pure that been discovered with assistance machine learning-demonstrating method by which learning can aid conjectures We propose process discover...
In recent years deep reinforcement learning (RL) systems have attained superhuman performance in a number of challenging task domains. However, major limitation such applications is their demand for massive amounts training data. A critical present objective thus to develop RL methods that can adapt rapidly new tasks. the work we introduce novel approach this challenge, which refer as meta-reinforcement learning. Previous has shown recurrent networks support meta-learning fully supervised...
State of the art deep reinforcement learning algorithms take many millions interactions to attain human-level performance. Humans, on other hand, can very quickly exploit highly rewarding nuances an environment upon first discovery. In brain, such rapid is thought depend hippocampus and its capacity for episodic memory. Here we investigate whether a simple model hippocampal control learn solve difficult sequential decision-making tasks. We demonstrate that it not only attains strategy...
In this work we explore a straightforward variational Bayes scheme for Recurrent Neural Networks. Firstly, show that simple adaptation of truncated backpropagation through time can yield good quality uncertainty estimates and superior regularisation at only small extra computational cost during training, also reducing the amount parameters by 80\%. Secondly, demonstrate how novel kind posterior approximation yields further improvements to performance Bayesian RNNs. We incorporate local...
Knowledge about social hierarchies organizes human behavior, yet we understand little the underlying computations. Here show that a Bayesian inference scheme, which tracks power of individuals, better captures behavioral and neural data compared with reinforcement learning model inspired by rating systems used in games such as chess. We provide evidence medial prefrontal cortex (MPFC) selectively mediates updating knowledge one's own hierarchy, opposed to another individual, process...
Atari games have been a long-standing benchmark in the reinforcement learning (RL) community for past decade. This was proposed to test general competency of RL algorithms. Previous work has achieved good average performance by doing outstandingly well on many set, but very poorly several most challenging games. We propose Agent57, first deep agent that outperforms standard human all 57 To achieve this result, we train neural network which parameterizes family policies ranging from...
Automated discovery of early visual concepts from raw image data is a major open challenge in AI research. Addressing this problem, we propose an unsupervised approach for learning disentangled representations the underlying factors variation. We draw inspiration neuroscience, and show how can be achieved generative model by applying same pressures as have been suggested to act ventral stream brain. By enforcing redundancy reduction, encouraging statistical independence, exposure with...