Google Inc.

Mark M Sandler

Andrew Gerald Howard

Andrey Zhmoginov

Dandapani Sivakumar

Menglong Zhu

Liang-Chieh Chen

Pramod Kaushik Mudrakarta

Ruoming Pang

Krzysztof Duleba

Kushal Dave

Luuk Van Dijk

Andras Erdei

Nir Ailon

Raoul Sam Daruwala

The present disclosure provides systems and methods that enable parameter-efficient transfer learning, multi-task learning, and/or other forms of model re-purposing such as model personalization or domain adaptation. In particular, as one example, a computing system can obtain a machine-learned model that has been previously trained on a first training dataset to perform a first task. The machine-learned model can include a first set of learnable parameters. The computing system can modify the machine-learned model to include a model patch, where the model patch includes a second set of learnable parameters. The computing system can train the machine-learned model on a second training dataset to perform a second task that is different from the first task, which may include learning new values for the second set of learnable parameters included in the model patch while keeping at least some (e.g., all) of the first set of parameters fixed.

Parameter-efficient multi-task and transfer learning

The present disclosure provides directed to new, more efficient neural network architectures. As one example, in some implementations, the neural network architectures of the present disclosure can include a linear bottleneck layer positioned structurally prior to and/or after one or more convolutional layers, such as, for example, one or more depthwise separable convolutional layers. As another example, in some implementations, the neural network architectures of the present disclosure can include one or more inverted residual blocks where the input and output of the inverted residual block are thin bottleneck layers, while an intermediate layer is an expanded representation. For example, the expanded representation can include one or more convolutional layers, such as, for example, one or more depthwise separable convolutional layers. A residual shortcut connection can exist between the thin bottleneck layers that play a role of an input and output of the inverted residual block.

Highly efficient convolutional neural networks

A computer-implemented method for identifying on-line comments as being legitimate or illegitimate is disclosed. The method includes receiving a comment directed to a piece of on-line content, randomly determining whether to review the comment manually, and providing for review of information regarding the comment by a manual reviewer if a determination is made to manually review the comment. The chance of determining whether to review the comment manually is dependent on outcomes of prior manual reviews of received comments.

Monitoring of negative feedback systems

Among other disclosed subject matter, a computer-implemented method includes receiving a first, a second and a third network address mapping. The method includes mapping each of the network addresses in the first, second and third network address mappings to respective first, second and third locations. The method includes determining, for each of the network addresses in the first, second and third network address mappings, location agreements. The method includes determining a first and second network address mapping agreement rate, a first and third network address mapping agreement rate, and a second and third network address mapping agreement rate, respectively based on location agreements determined for each of the network addresses. The method includes determining a confidence level associated with the first network address mapping based on the agreement rates.

Evaluating IP-location mapping data

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for scoring resources. One method includes generating a resource graph, the graph including nodes representing resources, nodes representing queries, edges between nodes representing resources, and edges between nodes representing queries and nodes representing resources. The method further includes calculating a score for each of the resources, wherein the score for a particular resource is based on scores for resources represented by nodes with edges to the node representing particular resource in the graph and scores for queries represented by nodes that have an edge to the node representing particular resource in the graph, and ranking one or more of the resources in response to a first query received from a user based, at least in part, on the scores.

Augmented resource graph for scoring resources

Methods, systems, and apparatus, including medium-encoded computer program products, for searching a data set and returning search results organized in a hierarchy of categories are disclosed. A set of categories is provided for organizing a set of search results, wherein each category is associated with one or more search results. The set of search results is organized into a hierarchy of categories, the hierarchy including at least one category from the set of categories. At least a portion of the hierarchy of categories is displayed and a user request to modify the hierarchy of categories is received. The hierarchy of categories is modified in accordance with the user request.

Growing community of inventors

Jersey City, NJ, United States of America

Mark M Sandler