Recommending content using neural networks

1. A method performed by one or more computers, the method comprising:

receiving context information for an action recommendation, wherein the action recommendation is a selection from a plurality of possible actions;

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, wherein each Bayesian neural network layer receives an input and applies a transformation to the input to generate an output from the received input in accordance with current values for each of a set of parameters, wherein each parameter for each Bayesian neural network layer is associated with data representing a probability distribution over a plurality of possible current values for the parameter, wherein the action score for a given action represents an expected return received if the action is selected, wherein the expected return is a combination of the expected reward received if the action is selected and a plurality of expected future rewards received in response to future action selections, and wherein processing the context information using the neural network to select the action comprises:

for each parameter of each Bayesian neural network layer, selecting the current value for the parameter using the data representing the probability distribution over the plurality of possible current values for the parameter; and

selecting an action from plurality of possible actions using the parameters of the distributions over the possible action scores for the action.

2. The method of claim 1 , wherein, for each of the parameters, the data representing the probability distribution over the plurality of possible current values is hyperparameters for the parameter, wherein the hyperparameters for the parameter are parameters of an approximation of the probability distribution for the parameter.

3. The method of claim 2 , further comprising:

training the neural network on training data to adjust the values of the hyperparameters of the Bayesian neural network layers.

4. The method of claim 3 , wherein training the neural network on training data comprises training the neural network using stochastic backpropagation.

5. The method of claim 1 , wherein the action score for a given action represents an expected reward received if the action is selected.

6. The method of claim 1 , further comprising:

receiving a reward in response to selecting the action;

generating a training observation that includes the context information, the selected action, and the received reward; and

storing the training observation as training data for use in training the neural network.

7. The method of claim 1 , further comprising:

receiving a reward in response to selecting the action;

generating a training observation that includes the context information, the selected action, and the received reward; and

training the neural network on-line using the training observation.

8. The method of claim 1 , wherein the one or more parameters of the distribution over possible action scores is an average of the possible action scores.

9. The method of claim 8 , wherein selecting an action score comprises selecting an action score having a highest average.

10. A computer program product encoded on one or more non-transitory computer storage media, the computer program product comprising instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

receiving context information for an action recommendation, wherein the action recommendation is a selection from a plurality of possible actions;

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, wherein each Bayesian neural network layer receives an input and applies a transformation to the input to generate an output from the received input in accordance with current values for each of a set of parameters, wherein each parameter for each Bayesian neural network layer is associated with data representing a probability distribution over a plurality of possible current values for the parameter, wherein the action score for a given action represents an expected return received if the action is selected, wherein the expected return is a combination of the expected reward received if the action is selected and a plurality of expected future rewards received in response to future action selections, and wherein processing the context information using the neural network to select the action comprises:

for each parameter of each Bayesian neural network layer, selecting the current value for the parameter using the data representing the probability distribution over the plurality of possible current values for the parameter; and

selecting an action from plurality of possible actions using the parameters of the distributions over the possible action scores for the action.

11. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising:

receiving context information for an action recommendation, wherein the action recommendation is a selection from a plurality of possible actions;

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, wherein each Bayesian neural network layer receives an input and applies a transformation to the input to generate an output from the received input in accordance with current values for each of a set of parameters, wherein each parameter for each Bayesian neural network layer is associated with data representing a probability distribution over a plurality of possible current values for the parameter, wherein the action score for a given action represents an expected return received if the action is selected, wherein the expected return is a combination of the expected reward received if the action is selected and a plurality of expected future rewards received in response to future action selections, and wherein processing the context information using the neural network to select the action comprises:

for each parameter of each Bayesian neural network layer, selecting the current value for the parameter using the data representing the probability distribution over the plurality of possible current values for the parameter; and

selecting an action from plurality of possible actions using the parameters of the distributions over the possible action scores for the action.

12. The system of claim 11 , wherein, for each of the parameters, the data representing the probability distribution over the plurality of possible current values is hyperparameters for the parameter, wherein the hyperparameters for the parameter are parameters of an approximation of the probability distribution for the parameter.

13. The system of claim 12 , further comprising:

training the neural network on training data to adjust the values of the hyperparameters of the Bayesian neural network layers.

14. The system of claim 11 , wherein the action score for a given action represents an expected reward received if the action is selected.

15. The system of claim 11 , further comprising:

receiving a reward in response to selecting the action;

generating a training observation that includes the context information, the selected action, and the received reward; and

storing the training observation as training data for use in training the neural network.

16. The system of claim 11 , further comprising:

receiving a reward in response to selecting the action;

generating a training observation that includes the context information, the selected action, and the received reward; and

training the neural network on-line using the training observation.

17. The system of claim 11 , wherein the one or more parameters of the distribution over possible action scores is an average of the possible action scores.

18. The system of claim 17 , wherein selecting an action score comprises selecting an action score having a highest average.