TECHNIQUES FOR TRAINING VISION FOUNDATION MODELS VIA MULTI-TEACHER DISTILLATION

    公开(公告)号:US20250165777A1

    公开(公告)日:2025-05-22

    申请号:US18740294

    申请日:2024-06-11

    Abstract: One embodiment of a method for training a first machine learning model includes processing first data via a plurality of trained machine learning models to generate a plurality of first outputs, processing the first data via the first machine learning model to generate a second output, processing the second output via a plurality of projection heads to generate a plurality of third outputs, computing a plurality of losses based on the plurality of first outputs and the plurality of third outputs, and performing one or more operations to update one or more parameters of the first machine learning model and one or more parameters of the plurality of projection heads based on the plurality of losses.

    MANAGING A SYSTEM LEVEL CACHE
    5.
    发明申请

    公开(公告)号:US20250156326A1

    公开(公告)日:2025-05-15

    申请号:US18505952

    申请日:2023-11-09

    Abstract: Various embodiments include techniques for managing cache memory in a computing system. The disclosed techniques include a cache policy manager that monitors activity of various components that access a common cache memory. The cache policy manager establishes cache rules that determine what data remains stored in cache memory and what data is removed from cache memory in order to make room for new data. As the activities of these components change over time, cache rules that work well for a previous activity profile may no longer work well for the current activity profile. Therefore, the cache policy manager dynamically modifies the cache rules as the activity profile changes in order to select cache rules at any given time that work well with the current activity profile. These techniques are advantageous over conventional approaches that employ static cache rules that work well only for specific activity profiles.

    NON-UNIFORM ALLOCATION OF SYMMETRIC MEMORY IN PARALLEL PROGRAMS

    公开(公告)号:US20250156314A1

    公开(公告)日:2025-05-15

    申请号:US18507659

    申请日:2023-11-13

    Abstract: Systems and methods herein are for at least one circuit to perform at least one process of different processes that may be associated with one or more applications, where the process may use one part of a virtual memory based on different requests by the different processes, where the virtual memory may include parts of equal allocations based on a maximum of different memory sizes in specifications associated with the different processes, where the one part of the virtual memory may be in a mapping with respect to one part of a physical memory of different allocated sizes, and where a translation for the mapping can occur using a start address and the maximum of the different memory sizes.

    HANG RECOVERY AND ERROR REPORTING ARCHITECTURE ON FPGA-BASED CONTROLLER SOLUTIONS

    公开(公告)号:US20250156260A1

    公开(公告)日:2025-05-15

    申请号:US19014072

    申请日:2025-01-08

    Abstract: A device includes a first finite state machine, and a controller coupled to the first finite state machine. The controller is to detect that the first finite state machine fails to perform a first operation within a first duration, and responsive to detecting that the first finite state machine fails to perform the first operation within the first duration, render the first finite state machine available to process a second operation by storing a reset value to a first register associated with the first finite state machine.

    History clamping for denoising dynamic ray-traced scenes using temporal accumulation

    公开(公告)号:US12299855B2

    公开(公告)日:2025-05-13

    申请号:US18480034

    申请日:2023-10-03

    Inventor: Pawel Kozlowski

    Abstract: Approaches presented herein can reduce temporal lag that may be introduced in a generated image sequence that utilizes temporal accumulation for denoising in dynamic scenes. A fast historical frame can be generated along with a full historical frame generated for a denoising process, with the fast historical frame being accumulated using an exponential moving average with a significantly higher blend weight. This fast history frame can be used to determine a clamping window that can be used to clamp a corresponding full historical value before, or after, reprojection. The fast historical blend weight can be adjusted to control the amount of noise versus temporal lag in an image sequence. In some embodiments, differences between fast and full historical values can also be used to determine an amount of spatial filtering to be applied.

    SYSTEMS AND METHODS FOR PERFORMING OPERATIONS IN A VEHICLE USING GAZE DETECTION

    公开(公告)号:US20250148810A1

    公开(公告)日:2025-05-08

    申请号:US19019008

    申请日:2025-01-13

    Abstract: In various examples, systems and methods are disclosed herein for a vehicle command operation system that may use technology across multiple modalities to cause vehicular operations to be performed in response to determining a focal point based on a gaze of an occupant. The system may utilize sensors to receive first data indicative of an eye gaze of an occupant of the vehicle. The system may utilize sensors to receive second data indicative of other data from the occupant. The system may then calculate a gaze vector based on the data indicative of the eye gaze of the occupant. The system may determine a focal point based on the gaze vector. In response to determining the focal point, the system causes an operation to be performed in the vehicle based on the second data.

Patent Agency Ranking