Google Inc.

Marcello Mathias Herreshoff

Jason Holt

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for constructing and processing computational graphs that represent dynamically structured machine learning models are disclosed. An example system receives data identifying a plurality of operations that can be performed on input data for processing by a dynamically structured machine learning model. The system also receives a plurality of labels corresponding to arguments for the plurality of operations. A directed computational graph representing a comprehensive layer of the dynamically structured machine learning model is generated from the identified operations and labels. An example system then receives an input for processing by the machine learning model and specifies data flow through the directed computational graph.

Constructing and processing computational graphs for dynamically structured machine learning models

The present disclosure provides systems and methods that enable adaptive training of a channel coding model including an encoder model, a channel model positioned structurally after the encoder model, and a decoder model positioned structurally after the channel model. The channel model can have been trained to emulate a communication channel, for example, by training the channel model on example data that has been transmitted via the communication channel. The channel coding model can be trained on a loss function that describes a difference between input data input into the encoder model and output data received from the decoder model. In particular, such a loss function can be backpropagated through the decoder model while modifying the decoder model, backpropagated through the channel model while the channel model is held constant, and then backpropagated through the encoder model while modifying the encoder model.

Adaptive channel coding using machine-learned models

The present disclosure provides systems and methods that enable training of an encoder model based on a decoder model that performs an inverse transformation relative to the encoder model. In one example, an encoder model can receive a first set of inputs and output a first set of outputs. The encoder model can be a neural network. The decoder model can receive the first set of outputs and output a second set of outputs. A loss function can describe a difference between the first set of inputs and the second set of outputs. According to an aspect of the present disclosure, the loss function can be sequentially backpropagated through the decoder model without modifying the decoder model and then through the encoder model while modifying the encoder model, thereby training the encoder model. Thus, an encoder model can be trained to have enforced consistency relative to the inverse decoder model.

Growing community of inventors

Mountain View, CA, United States of America

Marcello Mathias Herreshoff