Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for selecting answers to questions about documents. One of the methods includes receiving a document comprising a plurality of document tokens; receiving a question associated with the document, the question comprising a plurality of question tokens; processing the document tokens and the question tokens using a reader neural network to generate a joint numeric representation of the document and the question; and selecting, from the plurality of document tokens, an answer to the question using the joint numeric representation of the document and the question.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting actions to be performed by an agent interacting with an environment. In one aspect, a system comprises a grid cell neural network and an action selection neural network. The grid cell network is configured to: receive an input comprising data characterizing a velocity of the agent; process the input to generate a grid cell representation; and process the grid cell representation to generate an estimate of a position of the agent in the environment; the action selection neural network is configured to: receive an input comprising a grid cell representation and an observation characterizing a state of the environment; and process the input to generate an action selection network output.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing speech recognition by generating a neural network output from an audio data input sequence, where the neural network output characterizes words spoken in the audio data input sequence. One of the methods includes, for each of the audio data inputs, providing a current audio data input sequence that comprises the audio data input and the audio data inputs preceding the audio data input in the audio data input sequence to a convolutional subnetwork comprising a plurality of dilated convolutional neural network layers, wherein the convolutional subnetwork is configured to, for each of the plurality of audio data inputs: receive the current audio data input sequence for the audio data input, and process the current audio data input sequence to generate an alternative representation for the audio data input.

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: Systems, methods, and apparatus, including computer programs encoded on a computer storage medium, for selecting an actions from a set of actions to be performed by an agent interacting with an environment. In one aspect, the system includes a dueling deep neural network. The dueling deep neural network includes a value subnetwork, an advantage subnetwork, and a combining layer. The value subnetwork processes a representation of an observation to generate a value estimate. The advantage subnetwork processes the representation of the observation to generate an advantage estimate for each action in the set of actions. The combining layer combines the value estimate and the respective advantage estimate for each action to generate a respective Q value for the action. The system selects an action to be performed by the agent in response to the observation using the respective Q values for the actions in the set of actions.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a target sequence from an input sequence. In one aspect, a method comprises maintaining a set of current hypotheses, wherein each current hypothesis comprises an input prefix and an output prefix. For each possible combination of input and output prefix length, the method extends any current hypothesis that could reach the possible combination to generate respective extended hypotheses for each such current hypothesis; determines a respective direct score for each extended hypothesis using a direct model; determines a first number of highest-scoring hypotheses according to the direct scores; rescores the first number of highest-scoring hypotheses using a noisy channel model to generate a reduced number of hypotheses; and adds the reduced number of hypotheses to the set of current hypotheses.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for implementing grid Long Short-Term Memory (LSTM) neural networks that includes a plurality of N-LSTM blocks arranged in an N-dimensional grid. Each N-LSTM block is configured to: receive N input hidden vectors, the N input hidden vectors each corresponding to a respective one of the N dimensions; receive N input memory vectors, the N input memory vectors each corresponding to a respective one of the N dimensions; and, for each of the dimensions, apply a respective transform for the dimension to the memory hidden vector corresponding to the dimension and the input hidden vector corresponding to the dimension to generate a new hidden vector corresponding to the dimension and a new memory vector corresponding to the dimension.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for evaluating reinforcement learning policies. One of the methods includes receiving a plurality of training histories for a reinforcement learning agent; determining a total reward for each training observation in the training histories; partitioning the training observations into a plurality of partitions; determining, for each partition and from the partitioned training observations, a probability that the reinforcement learning agent will receive the total reward for the partition if the reinforcement learning agent performs the action for the partition in response to receiving the current observation; determining, from the probabilities and for each total reward, a respective estimated value of performing each action in response to receiving the current observation; and selecting an action from the pre-determined set of actions from the estimated values in accordance with an action selection policy.

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 computer storage media, for content recommendation using neural networks. One of the methods includes receiving context information for an action recommendation; processing the context information using a neural network that comprises one or more Bayesian neural network layers to generate, for each of the actions, one or more parameters of a distribution over possible action scores for the action and selecting an action from plurality of possible actions using the parameters of the distributions over the possible action scores for the action.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for compressing images using neural networks. One of the methods includes receiving an image; processing the image using an encoder neural network, wherein the encoder neural network is configured to receive the image and to process the image to generate an output defining values of a first number of latent variables that each represent a feature of the image; generating a compressed representation of the image using the output defining the values of the first number of latent variables; and providing the compressed representation of the image for use in generating a reconstruction of the image.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for augmenting neural networks with an external memory. One of the methods includes providing an output derived from the neural network output for the time step as a system output for the time step; maintaining a current state of the external memory; determining, from the neural network output for the time step, memory state parameters for the time step; updating the current state of the external memory using the memory state parameters for the time step; reading data from the external memory in accordance with the updated state of the external memory; and combining the data read from the external memory with a system input for the next time step to generate the neural network input for the next time step.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating images using neural networks. One of the methods includes generating the output image pixel by pixel from a sequence of pixels taken from the output image, comprising, for each pixel in the output image, generating a respective score distribution over a discrete set of possible color values for each of the plurality of color channels.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a variational auto-encoder (VAE) to generate disentangled latent factors on unlabeled training images. In one aspect, a method includes receiving the plurality of unlabeled training images, and, for each unlabeled training image, processing the unlabeled training image using the VAE to determine the latent representation of the unlabeled training image and to generate a reconstruction of the unlabeled training image in accordance with current values of the parameters of the VAE, and adjusting current values of the parameters of the VAE by optimizing a loss function that depends on a quality of the reconstruction and also on a degree of independence between the latent factors in the latent representation of the unlabeled training image.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for neural machine translation. In one aspect, a system is configured to receive an input sequence of source embeddings representing a source sequence of words in a source natural language and to generate an output sequence of target embeddings representing a target sequence of words that is a translation of the source sequence into a target natural language, the system comprising: a dilated convolutional neural network configured to process the input sequence of source embeddings to generate an encoded representation of the source sequence, and a masked dilated convolutional neural network configured to process the encoded representation of the source sequence to generate the output sequence of target embeddings.

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 a computer storage medium, for training a recurrent neural network on training sequences using backpropagation through time. In one aspect, a method includes receiving a training sequence including a respective input at each of a number of time steps; obtaining data defining an amount of memory allocated to storing forward propagation information for use during backpropagation; determining, from the number of time steps in the training sequence and from the amount of memory allocated to storing the forward propagation information, a training policy for processing the training sequence, wherein the training policy defines when to store forward propagation information during forward propagation of the training sequence; and training the recurrent neural network on the training sequence in accordance with the training policy.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating an output sequence of audio data that comprises a respective audio sample at each of a plurality of time steps. 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: Systems, methods, and apparatus, including computer programs encoded on a computer storage medium, for selecting an actions from a set of actions to be performed by an agent interacting with an environment. In one aspect, the system includes a dueling deep neural network. The dueling deep neural network includes a value subnetwork, an advantage subnetwork, and a combining layer. The value subnetwork processes a representation of an observation to generate a value estimate. The advantage subnetwork processes the representation of the observation to generate an advantage estimate for each action in the set of actions. The combining layer combines the value estimate and the respective advantage estimate for each action to generate a respective Q value for the action. The system selects an action to be performed by the agent in response to the observation using the respective Q values for the actions in the set of actions.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a neural network used to select actions performed by a reinforcement learning agent interacting with an environment. In one aspect, a method includes maintaining a replay memory, where the replay memory stores pieces of experience data generated as a result of the reinforcement learning agent interacting with the environment. Each piece of experience data is associated with a respective expected learning progress measure that is a measure of an expected amount of progress made in the training of the neural network if the neural network is trained on the piece of experience data. The method further includes selecting a piece of experience data from the replay memory by prioritizing for selection pieces of experience data having relatively higher expected learning progress measures and training the neural network on the selected piece of experience data.