Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for a system configured to select actions to be performed by an agent that interacts with an environment. The system comprises a manager neural network subsystem and a worker neural network subsystem. The manager subsystem is configured to, at each of the multiple time steps, generate a final goal vector for the time step. The worker subsystem is configured to, at each of multiple time steps, use the final goal vector generated by the manager subsystem to generate a respective action score for each action in a predetermined set of actions.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a reinforcement learning system. In one aspect, a method of training an action selection policy neural network for use in selecting actions to be performed by an agent navigating through an environment to accomplish one or more goals comprises: receiving an observation image characterizing a current state of the environment; processing, using the action selection policy neural network, an input comprising the observation image to generate an action selection output; processing, using a geometry-prediction neural network, an intermediate output generated by the action selection policy neural network to predict a value of a feature of a geometry of the environment when in the current state; and backpropagating a gradient of a geometry-based auxiliary loss into the action selection policy neural network to determine a geometry-based auxiliary update for current values of the network parameters.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for distributed training of reinforcement learning systems. One of the methods includes receiving, by a learner, current values of the parameters of the Q network from a parameter server, wherein each learner maintains a respective learner Q network replica and a respective target Q network replica; updating, by the learner, the parameters of the learner Q network replica maintained by the learner using the current values; selecting, by the learner, an experience tuple from a respective replay memory; computing, by the learner, a gradient from the experience tuple using the learner Q network replica maintained by the learner and the target Q network replica maintained by the learner; and providing, by the learner, the computed gradient to the parameter server.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a reinforcement learning system. In one aspect, a method of training an action selection policy neural network for use in selecting actions to be performed by an agent navigating through an environment to accomplish one or more goals comprises: receiving an observation image characterizing a current state of the environment; processing, using the action selection policy neural network, an input comprising the observation image to generate an action selection output; processing, using a geometry-prediction neural network, an intermediate output generated by the action selection policy neural network to predict a value of a feature of a geometry of the environment when in the current state; and backpropagating a gradient of a geometry-based auxiliary loss into the action selection policy neural network to determine a geometry-based auxiliary update for current values of the network parameters.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for a system configured to select actions to be performed by an agent that interacts with an environment. The system comprises a manager neural network subsystem and a worker neural network subsystem. The manager subsystem is configured to, at each of the multiple time steps, generate a final goal vector for the time step. The worker subsystem is configured to, at each of multiple time steps, use the final goal vector generated by the manager subsystem to generate a respective action score for each action in a predetermined set of actions.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a machine learning model. In one aspect, a method includes receiving training data for training the machine learning model on a plurality of tasks, where each task includes multiple batches of training data. A task is selected in accordance with a current task selection policy. A batch of training data is selected from the selected task. The machine learning model is trained on the selected batch of training data to determine updated values of the model parameters. A learning progress measure that represents a progress of the training of the machine learning model as a result of training the machine learning model on the selected batch of training data is determined. The current task selection policy is updated using the learning progress measure.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for distributed training of reinforcement learning systems. One of the methods includes receiving, by a learner, current values of the parameters of the Q network from a parameter server, wherein each learner maintains a respective learner Q network replica and a respective target Q network replica; updating, by the learner, the parameters of the learner Q network replica maintained by the learner using the current values; selecting, by the learner, an experience tuple from a respective replay memory; computing, by the learner, a gradient from the experience tuple using the learner Q network replica maintained by the learner and the target Q network replica maintained by the learner; and providing, by the learner, the computed gradient to the parameter server.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a reinforcement learning system. In one aspect, a method of training an action selection policy neural network for use in selecting actions to be performed by an agent navigating through an environment to accomplish one or more goals comprises: receiving an observation image characterizing a current state of the environment; processing, using the action selection policy neural network, an input comprising the observation image to generate an action selection output; processing, using a geometry-prediction neural network, an intermediate output generated by the action selection policy neural network to predict a value of a feature of a geometry of the environment when in the current state; and backpropagating a gradient of a geometry-based auxiliary loss into the action selection policy neural network to determine a geometry-based auxiliary update for current values of the network parameters.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a reinforcement learning system. The method includes: training an action selection policy neural network, and during the training of the action selection neural network, training one or more auxiliary control neural networks and a reward prediction neural network. Each of the auxiliary control neural networks is configured to receive a respective intermediate output generated by the action selection policy neural network and generate a policy output for a corresponding auxiliary control task. The reward prediction neural network is configured to receive one or more intermediate outputs generated by the action selection policy neural network and generate a corresponding predicted reward.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for asynchronous deep reinforcement learning. One of the systems includes a plurality of workers, wherein each worker is configured to operate independently of each other worker, and wherein each worker is associated with a respective actor that interacts with a respective replica of the environment during the training of the deep neural network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for asynchronous deep reinforcement learning. One of the systems includes a plurality of workers, wherein each worker is configured to operate independently of each other worker, and wherein each worker is associated with a respective actor that interacts with a respective replica of the environment during the training of the deep neural network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing images using recurrent attention. One of the methods includes determining a location in the first image; extracting a glimpse from the first image using the location; generating a glimpse representation of the extracted glimpse; processing the glimpse representation using a recurrent neural network to update a current internal state of the recurrent neural network to generate a new internal state; processing the new internal state to select a location in a next image in the image sequence after the first image; and processing the new internal state to select an action from a predetermined set of possible actions.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing inputs using an image processing neural network system that includes a spatial transformer module. One of the methods includes receiving an input feature map derived from the one or more input images, and applying a spatial transformation to the input feature map to generate a transformed feature map, comprising: processing the input feature map to generate spatial transformation parameters for the spatial transformation, and sampling from the input feature map in accordance with the spatial transformation parameters to generate the transformed feature map.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for asynchronous deep reinforcement learning. One of the systems includes a plurality of workers, wherein each worker is configured to operate independently of each other worker, and wherein each worker is associated with a respective actor that interacts with a respective replica of the environment during the training of the deep neural network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing inputs using an image processing neural network system that includes a spatial transformer module. One of the methods includes receiving an input feature map derived from the one or more input images, and applying a spatial transformation to the input feature map to generate a transformed feature map, comprising: processing the input feature map to generate spatial transformation parameters for the spatial transformation, and sampling from the input feature map in accordance with the spatial transformation parameters to generate the transformed feature map.

Abstract: Methods and systems for performing a sequence of machine learning tasks. One system includes a sequence of deep neural networks (DNNs), including: a first DNN corresponding to a first machine learning task, wherein the first DNN comprises a first plurality of indexed layers, and each layer in the first plurality of indexed layers is configured to receive a respective layer input and process the layer input to generate a respective layer output; and one or more subsequent DNNs corresponding to one or more respective machine learning tasks, wherein each subsequent DNN comprises a respective plurality of indexed layers, and each layer in a respective plurality of indexed layers with index greater than one receives input from a preceding layer of the respective subsequent DNN, and one or more preceding layers of respective preceding DNNs, wherein a preceding layer is a layer whose index is one less than the current index.

Abstract: We describe a method of reinforcement learning for a subject system having multiple states and actions to move from one state to the next. Training data is generated by operating on the system with a succession of actions and used to train a second neural network. Target values for training the second neural network are derived from a first neural network which is generated by copying weights of the second neural network at intervals.

Abstract: We describe a method of reinforcement learning for a subject system having multiple states and actions to move from one state to the next. Training data is generated by operating on the system with a succession of actions and used to train a second neural network. Target values for training the second neural network are derived from a first neural network which is generated by copying weights of the second neural network at intervals.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for asynchronous deep reinforcement learning. One of the systems includes a plurality of workers, wherein each worker is configured to operate independently of each other worker, and wherein each worker is associated with a respective actor that interacts with a respective replica of the environment during the training of the deep neural network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing inputs using an image processing neural network system that includes a spatial transformer module. One of the methods includes receiving an input feature map derived from the one or more input images, and applying a spatial transformation to the input feature map to generate a transformed feature map, comprising: processing the input feature map to generate spatial transformation parameters for the spatial transformation, and sampling from the input feature map in accordance with the spatial transformation parameters to generate the transformed feature map.