Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating sequences of predicted observations, for example images. In one aspect, a system comprises a controller recurrent neural network, and a decoder neural network to process a set of latent variables to generate an observation. An external memory and a memory interface subsystem is configured to, for each of a plurality of time steps, receive an updated hidden state from the controller, generate a memory context vector by reading data from the external memory using the updated hidden state, determine a set of latent variables from the memory context vector, generate a predicted observation by providing the set of latent variables to the decoder neural network, write data to the external memory using the latent variables, the updated hidden state, or both, and generate a controller input for a subsequent time step from the latent variables.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting an action to be performed by a reinforcement learning agent. The method includes obtaining an observation characterizing a current state of an environment. For each layer parameter of each noisy layer of a neural network, a respective noise value is determined. For each layer parameter of each noisy layer, a noisy current value for the layer parameter is determined from a current value of the layer parameter, a current value of a corresponding noise parameter, and the noise value. A network input including the observation is processed using the neural network in accordance with the noisy current values to generate a network output for the network input. An action is selected from a set of possible actions to be performed by the agent in response to the observation using the network output.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for executing a signal generation neural network on parallel processing hardware. One of the methods includes receiving weight matrices of a layer of a signal generation neural network. Rows of a first matrix for the layer are interleaved by assigning groups of rows of the first matrix to respective thread blocks of a plurality of thread blocks. A first subset of rows of the one or more other weight matrices are assigned to a first subset of the plurality of thread blocks and a second subset of rows of the one or more other weight matrices are assigned to a second subset of the plurality of thread blocks. The first matrix operation is performed substantially in parallel by the plurality of thread blocks. The other matrix operations are performed substantially in parallel by the plurality of thread blocks.

Abstract: A system comprising a causal convolutional neural network to autoregressively generate a succession of values of a data item conditioned upon previously generated values of the data item. The system includes support memory for a set of support data patches each of which comprises an encoding of an example data item. A soft attention mechanism attends to one or more patches when generating the current item value. The soft attention mechanism determines a set of scores for the support data patches, for example in the form of a soft attention query vector dependent upon the previously generated values of the data item. The soft attention query vector is used to query the memory. When generating the value of the data item at a current iteration layers of the causal convolutional neural network are conditioned upon the support data patches weighted by the scores.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for parallel processing of video frames using neural networks. One of the methods includes receiving a video sequence comprising a respective video frame at each of a plurality of time steps; and processing the video sequence using a video processing neural network to generate a video processing output for the video sequence, wherein the video processing neural network includes a sequence of network components, wherein the network components comprise a plurality of layer blocks each comprising one or more neural network layers, wherein each component is active for a respective subset of the plurality of time steps, and wherein each layer block is configured to, at each time step at which the layer block is active, receive an input generated at a previous time step and to process the input to generate a block output.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training an action selection neural network that is used to select actions to be performed by a reinforcement learning agent interacting with an environment. In particular, the actions are selected from a continuous action space and the system trains the action selection neural network jointly with a distribution Q network that is used to update the parameters of the action selection neural network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating discrete latent representations of input data items. One of the methods includes receiving an input data item; providing the input data item as input to an encoder neural network to obtain an encoder output for the input data item; and generating a discrete latent representation of the input data item from the encoder output, comprising: for each of the latent variables, determining, from a set of latent embedding vectors in the memory, a latent embedding vector that is nearest to the encoded vector for the latent variable.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing sequences using convolutional neural networks. One of the methods includes, for each of the time steps: providing a current sequence of audio data as input to a convolutional subnetwork, wherein the current sequence comprises the respective audio sample at each time step that precedes the time step in the output sequence, and wherein the convolutional subnetwork is configured to process the current sequence of audio data to generate an alternative representation for the time step; and providing the alternative representation for the time step as input to an output layer, wherein the output layer is configured to: process the alternative representation to generate an output that defines a score distribution over a plurality of possible audio samples for the time step.

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 training a neural network. A method includes: training a neural network having a plurality of network parameters to perform a particular neural network task and to determine trained values of the network parameters using an iterative training process having a plurality of hyperparameters, the method comprising: maintaining a plurality of candidate neural networks and, for each of the candidate neural networks, data specifying: (i) respective values of the network parameters for the candidate neural network, (ii) respective values of the hyperparameters for the candidate neural network, and (iii) a quality measure that measures a performance of the candidate neural network on the particular neural network task; and for each of the plurality of candidate neural networks, repeatedly performing additional training operations.

Abstract: A feedforward generative neural network that generates an output example that includes multiple output samples of a particular type in a single neural network inference. Optionally, the generation may be conditioned on a context input. For example, the feedforward generative neural network may generate a speech waveform that is a verbalization of an input text segment conditioned on linguistic features of the text segment.

Abstract: Techniques are disclosed that enable generation of an audio waveform representing synthesized speech based on a difference signal determined using an autoregressive model. Various implementations include using a distribution of the difference signal values to represent sounds found in human speech with a higher level of granularity than sounds not frequently found in human speech. Additional or alternative implementations include using one or more speakers of a client device to render the generated audio waveform.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a machine learning model. A method includes: maintaining a plurality of training sessions; assigning, to each worker of one or more workers, a respective training session of the plurality of training sessions; repeatedly performing operations until meeting one or more termination criteria, the operations comprising: receiving an updated training session from a respective worker of the one or more workers, selecting a second training session, selecting, based on comparing the updated training session and the second training session using a fitness evaluation function, either the updated training session or the second training session as a parent training session, generating a child training session from the selected parent training session, and assigning the child training session to an available worker, and selecting a candidate model to be a trained model for the machine learning model.

Abstract: A computer-implemented method for automatically determining a neural network architecture represents a neural network architecture as a data structure defining a hierarchical set of directed acyclic graphs in multiple levels. Each graph has an input, an output, and a plurality of nodes between the input and the output. At each level, a corresponding set of the nodes are connected pairwise by directed edges which indicate operations performed on outputs of one node to generate an input to another node. Each level is associated with a corresponding set of operations. At a lowest level, the operations associated with each edge are selected from a set of primitive operations. The method includes repeatedly generating new sample neural network architectures, and evaluating their fitness. The modification is performed by selecting a level, selecting two nodes at that level, and modifying, removing or adding an edge between those nodes according to operations associated with lower levels of the hierarchy.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for selecting actions from large discrete action sets. One of the methods includes receiving a particular observation representing a particular state of an environment; and selecting an action from a discrete set of actions to be performed by an agent interacting with the environment, comprising: processing the particular observation using an actor policy network to generate an ideal point; determining, from the points that represent actions in the set, the k nearest points to the ideal point; for each nearest point of the k nearest points: processing the nearest point and the particular observation using a Q network to generate a respective Q value for the action represented by the nearest point; and selecting the action to be performed by the agent from the k actions represented by the k nearest points based on the Q values.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for reinforcement learning with scheduled auxiliary tasks. In one aspect, a method includes maintaining data specifying parameter values for a primary policy neural network and one or more auxiliary neural networks; at each of a plurality of selection time steps during a training episode comprising a plurality of time steps: receiving an observation, selecting a current task for the selection time step using a task scheduling policy, processing an input comprising the observation using the policy neural network corresponding to the selected current task to select an action to be performed by the agent in response to the observation, and causing the agent to perform the selected action.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a Q network used to select actions to be performed by an agent interacting with an environment. One of the methods includes obtaining a plurality of experience tuples and training the Q network on each of the experience tuples using the Q network and a target Q network that is identical to the Q network but with the current values of the parameters of the target Q network being different from the current values of the parameters of the Q network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting an action to be performed by a reinforcement learning agent interacting with an environment. In one aspect, a method comprises: receiving a current observation; for each action of a plurality of actions: randomly sampling one or more probability values; for each probability value: processing the action, the current observation, and the probability value using a quantile function network to generate an estimated quantile value for the probability value with respect to a probability distribution over possible returns that would result from the agent performing the action in response to the current observation; determining a measure of central tendency of the one or more estimated quantile values; and selecting an action to be performed by the agent in response to the current observation using the measures of central tendency for the actions.

Abstract: An off-policy reinforcement learning actor-critic neural network system configured to select actions from a continuous action space to be performed by an agent interacting with an environment to perform a task. An observation defines environment state data and reward data. The system has an actor neural network which learns a policy function mapping the state data to action data. A critic neural network learns an action-value (Q) function. A replay buffer stores tuples of the state data, the action data, the reward data and new state data. The replay buffer also includes demonstration transition data comprising a set of the tuples from a demonstration of the task within the environment. The neural network system is configured to train the actor neural network and the critic neural network off-policy using stored tuples from the replay buffer comprising tuples both from operation of the system and from the demonstration transition data.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a generator neural network and an encoder neural network. The generator neural network generates, based on a set of latent values, data items which are samples of a distribution. The encoder neural network generates a set of latent values for a respective data item. The training method comprises jointly training the generator neural network, the encoder neural network and a discriminator neural network configured to distinguish between samples generated by the generator network and samples of the distribution which are not generated by the generator network. The discriminator neural network is configured to distinguish by processing, by the discriminator neural network, an input pair comprising a sample part and a latent part.