University of Southern California

Gm Global Technolgoy Operations, Inc.

Coinbase, Inc.

Bhaskar Krishnamachari

Fan Bai

Suvil Singh Deora

Majed Alresaini

Michael J Neely

Rajarshi Gupta

Kwame-Lante Wright

Vijay Dialani

In certain embodiments, model unlearning verifiability may be facilitated. In some embodiments, in connection with a request for unlearning of a dataset for a machine learning model, an unlearning process may be performed on the machine learning model, where the unlearning process involves multiple stages, for the unlearning of the dataset, at which model instances of the machine learning model are respectively generated. One or more proofs or hash-based values corresponding to the model instances may be generated, such as first and second hash-based values corresponding to first and second model instances of the machine learning model generated at first and second stages of the unlearning process, respectively (e.g., where the second stage depends on the first stage). The proofs or hash-based values may then be stored at one or more database locations of a database to enable verification of the performance of the unlearning process.

Systems and methods for facilitating verifiability of machine learning model unlearning

A task allocation system for dynamically allocating computing tasks to networked computing resources with heterogeneous capabilities includes: a task library having a plurality of tasks, each task being one of a plurality of different implementations of an application, wherein the different implementations of an application provide different levels of accuracy and resource usage during execution, wherein the different implementations are configured based on a trade-off between level of accuracy and resource usage during execution; and a real-time scheduler module configured to monitor available computing resources and connectivity to the computing resources, receive a plurality of task requests, prioritize applications to be executed when performing the tasks wherein more critical applications are assigned a higher priority, allocate computing resources to the higher priority applications, allocate remaining computing resources to other applications, and select specific implementations of applications for assignment to specific computing resources in a way that maximizes an overall utility.

Method and apparatus of allocating automotive computing tasks to networked devices with heterogeneous capabilities

A method for localizing a mobile device in a physical space based on radio signals received from transmitters in the physical space includes a step of transforming a received signal strength vector to produce a transformed received signal strength vector. Sequence based localization is performed on the transformed RSS vector with a different ideal sequence centroid table.

Received signal strength transformation method to achieve better indoor radio frequency localization

Routing technologies are disclosed. A described technology includes routing packets through a network based on backpressure scheduling; when a number of packets in a transmitter queue satisfies a first threshold, retaining copies of at least a portion of the packets; and once the number of packets in the transmitter queue satisfies a second threshold, transmitting the retained copies of the packets to perform the routing. Retaining the copies can include copying a packet from the transmitter queue into a duplicate buffer when the number of packets in the transmitter queue is below the first threshold.

Growing community of inventors

Los Angeles, CA, United States of America

Bhaskar Krishnamachari