Methods and systems for relaying feature-driven communications
AssigneeHuawei Technologies Co., Ltd.
InventorsWuxian Shi, Yiqun Ge, Wen Tong, Qifan Zhang
Life of this patent
Estimating the term…
Abstract
Methods and apparatuses for feature-driven communications are described. A set of features describing an observed subject is transmitted by a transmitting electronic device (ED) to a base station (BS). The BS translates the received features to another set of transmission features to be transmitted to a receiving ED. The receiving ED recovers information about the subject from the features received from the BS.
Claims17
An apparatus for feature-based communications, the apparatus comprising:
- a receiver for receiving, from at least one transmitting electronic device (ED), a set of received features representing a subject, each received feature providing information about a respective probability distribution that each represents one or more aspects of the subject;
- a translator, implemented using a deep neural network (DNN), for encoding the set of received features into a set of machine-learning based transmission features representing the subject, each transmission feature providing information about a respective probability distribution that each represents one or more aspects of the subject, at least one probability distribution represented by the set of transmission features being different from the set of received features;
- a reducer configured to reduce the set of transmission features;
- and
- a transmitter for transmitting the set of transmission features after reducing by the reducer to at least one receiving ED.
An apparatus for feature-based communications, the apparatus comprising:
- a receiver for receiving over a transmission channel, from a base station (BS), a set of machine-learning based transmitted features representing a subject, each transmitted feature providing information about a respective probability distribution that each represents one or more aspects of the subject;
- and
- a feature decoder DNN configured to decode, from the set of transmitted features, a set of recovered information about the subject, wherein the feature decoder DNN is trained to satisfy:
- a first target of maximizing likelihood between the set of recovered information and source information about the subject, and a second target of minimizing an information bottleneck to be within a predetermined physical channel capacity limit of the transmission channel.
A method for training an encoder and decoder pair for feature-based communications, the method comprising:
- receiving a set of training data samples, the training data samples representing probability distributions representing one or more features of a subject;
- training a probabilistic feature encoder and probabilistic feature decoder pair together, using joint optimization, the feature encoder and feature decoder pair being trained together to satisfy:
- a first target of maximizing likelihood between information recovered by the decoder and source information encoded by the encoder, and a second target of minimizing an upper boundary of mutual information to be within a predetermined physical channel capacity limit;
- and
- after determining that training has converged to satisfy the first and second targets, storing trained weights for the feature encoder and feature decoder pair.
Dependencies are inferred from the claim text.