Microsoft Technology Licensing, LLC

Lihong Li

Li Deng

Jianfeng Gao

Xiujun Li

Imed Zitouni

Jianshu Chen

Ji He

Faisal Ahmed

Jinyoung Kim

Pushmeet Kohli

Xiaodong He

Rishabh Singh

Dengyong Zhou

Fethiye Asli Celikyilmaz

John C Langford

Siddhartha Sen

Leon Bottou

Prabhdeep Singh

Aleksandrs Slivkins

Chong Wang

Emilio Parisotto

Miroslav Dudik

Bhuwan Dhingra

Yun-Nung Vivian Chen

Alekh Agarwal

Abdelrahman S A Mohamed

Mari Ostendorf

Sarah L Bird

Xiadong He

Zachary Chase Lipton

An example system includes a processor to receive production log data and pre-process the production log data to generate an exploration data set. The production log data includes continuous updates of a ranking function associated with different actions for a same query issued by a same user in different impressions. The processor can perform offline estimation of online metrics for ranking functions using the exploration data set, evaluate a proposed ranking function by comparing the estimated online metrics to a set of baseline metrics of a baseline ranking function, detect that the proposed ranking function is a preferred ranking function having a higher detected metric score than the baseline ranking function based on a comparison of the estimated online metrics with the baseline metrics of the baseline ranking function, and execute an action on an information retrieval system based on the preferred ranking function.

Offline evaluation of ranking functions

Systems and methods are disclosed for inquiry-based deep learning. In one implementation, a first content segment is selected from a body of content. The content segment includes a first content element. The first content segment is compared to a second content segment to identify a content element present in the first content segment that is not present in the second content segment. Based on an identification of the content element present in the first content segment that is not present in the second content segment, the content element is stored in a session memory. A first question is generated based on the first content segment. The session memory is processed to compute an answer to the first question. An action is initiated based on the answer. Using deep learning, content segments can be encoded into memory. Incremental questioning can serve to focus various deep learning operations on certain content segments.

Inquiry-based deep learning

The claimed subject matter includes techniques for offline evaluation of ranking functions. An example system includes a first module configured to receive production log data, the first module to pre-process the production log data to generate an exploration data set. The example system also includes a second module configured to perform offline estimation of online metrics for ranking functions using the exploration data set. The example system also includes a third module to evaluate a proposed ranking function by comparing the estimated online metrics to a set of baseline metrics of a baseline ranking function and detecting that the estimated online metrics of the proposed ranking function exceed, are lower than, or are within a predetermined range of the baseline metrics.

A processing unit can operate a first recurrent computational model (RCM) to provide first state information and a predicted result value. The processing unit can operating a first network computational model (NCM) to provide respective expectation values of a plurality of actions based at least in part on the first state information. The processing unit can provide an indication of at least one of the plurality of actions, and receive a reference result value, e.g., via a communications interface. The processing unit can train the first RCM based at least in part on the predicted result value and the reference result value to provide a second RCM, and can train the first NCM based at least in part on the first state information and the at least one of the plurality of actions to provide a second NCM.

Multi-model controller

A processing unit can determine a first feature value corresponding to a session by operating a first network computational model (NCM) based part on information of the session. The processing unit can determine respective second feature values corresponding to individual actions of a plurality of actions by operating a second NCM. The second NCM can use a common set of parameters in determining the second feature values. The processing unit can determine respective expectation values of some of the actions of the plurality of actions based on the first feature value and the respective second feature values. The processing unit can select a first action of the plurality of actions based on at least one of the expectation values. In some examples, the processing unit can operate an NCM to determine expectation values based on information of a session and information of respective actions.

Multiple-action computational model training and operation

Described are systems, methods, and computer-readable media for program generation in a domain-specific language based on input-output examples. In accordance with various embodiments, a neural-network-based program generation model conditioned on an encoded set of input-output examples is used to generate a program tree by iteratively expanding a partial program tree, beginning with a root node and ending when all leaf nodes are terminal.

Neural network for program synthesis

Described herein are systems, methods, and techniques by which a processing unit can build an end-to-end dialogue agent model for end-to-end learning of dialogue agents for information access and apply the end-to-end dialogue agent model with soft attention over knowledge base entries to make the dialogue system differentiable. In various examples the processing unit can apply the end-to-end dialogue agent model to a source of input, fill slots for output from the knowledge base entries, induce a posterior distribution over the entities in a knowledge base or induce a posterior distribution of a target of the requesting user over entities from a knowledge base, develop an end-to-end differentiable model of a dialogue agent, use supervised and/or imitation learning to initialize network parameters, calculate a modified version of an episodic algorithm. e.g., the REINFORCE algorithm, for training an end-to-end differentiable model based on user feedback.

End-to-end learning of dialogue agents for information access

The claimed subject matter includes techniques for providing an application testing service. An example method includes receiving context information from a client system, the context information comprising parameters that describe details of a user interaction with an application under test (AUT). The method also includes receiving a set of potential actions from the client system. The method also includes identifying a selected action from the set of potential actions and sending the selected action to the client system, wherein the AUT is customized based on the selected action. The method also includes receiving reward data from the client system based on a user's interaction with the AUT. The method also includes storing the context information, the selected action, and the reward data to a log of application test data.

Application testing

Efficient exploration of natural language conversations associated with dialogue policy learning may be performed using probabilistic distributions. Exploration may comprise identifying key terms associated with the received natural language input utilizing the structured representation. Identifying key terms may include converting raw text of the received natural language input into a structured representation. Exploration may also comprise mapping at least one of the key terms to an action to be performed by the computer system in response to receiving natural language input associated with the at least one key term. Mapping may then be performed using a probabilistic distribution. The action may then be performed by the computer system. A replay buffer may also be utilized by the computer system to track what has occurred in previous conversations. The replay buffer may then be pre-filled with one or more successful dialogues to jumpstart exploration.

Efficient dialogue policy learning

Growing community of inventors

Redmond, WA, United States of America

Lihong Li