Memory augmented generative temporal models

A method performed by one or more data processing apparatus, the method comprising, for each of a plurality of time steps:

• receiving a controller neural network input for the time step, wherein the controller neural network input for the time step comprises a set of latent variables generated at a previous time step that collectively define a compressed representation of a predicted observation characterizing an environment at the previous time step;
• processing the controller neural network input for the time step using a controller recurrent neural network, in accordance with a current hidden state of the controller recurrent neural network, to generate an updated hidden state of the controller recurrent neural network for the time step;
• generating a memory context vector by reading data from an external memory using the updated hidden state of the controller recurrent neural network for the time step;
• determining, from at least the memory context vector, a set of latent variables for the time step that collectively define a compressed representation of a predicted observation characterizing the environment at the time step;
• generating the predicted observation characterizing the environment at the time step, comprising:
• receiving a decoder neural network input comprising the set of latent variables;
• and
• processing the decoder neural network input comprising the set of latent variables to generate the predicted observation characterizing the environment at the time step;
• writing data to the external memory using the set of latent variables, the updated hidden state, or both;
• and
• generating a controller neural network input for a subsequent time step from the set of latent variables for the time step, wherein the controller neural network input for the subsequent time step comprises the set of latent variables for the time step.

One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising, for each of a plurality of time steps:

• receiving a controller neural network input for the time step, wherein the controller neural network input for the time step comprises a set of latent variables generated at a previous time step that collectively define a compressed representation of a predicted observation characterizing an environment at the previous time step;
• processing the controller neural network input for the time step using a controller recurrent neural network, in accordance with a current hidden state of the controller recurrent neural network, to generate an updated hidden state of the controller recurrent neural network for the time step;
• generating a memory context vector by reading data from an external memory using the updated hidden state of the controller recurrent neural network for the time step;
• determining, from at least the memory context vector, a set of latent variables for the time step that collectively define a compressed representation of a predicted observation characterizing the environment at the time step;
• generating the predicted observation characterizing the environment at the time step, comprising:
• receiving a decoder neural network input comprising the set of latent variables;
• and
• processing the decoder neural network input comprising the set of latent variables to generate the predicted observation characterizing the environment at the time step;
• writing data to the external memory using the set of latent variables, the updated hidden state, or both;
• and
• generating a controller neural network input for a subsequent time step from the set of latent variables for the time step, wherein the controller neural network input for the subsequent time step comprises the set of latent variables for the time step.