Meta Platforms Technologies, LLC

Barbara De Salvo

Syed Shakib Sarwar

Xinqiao Liu

Ziyun Li

Manan Suri

Hans Reyserhove

Asif Khan

Hsien-Hsin Sean Lee

Vivek Kamalkant Parmar

An artificial reality system performs local user adaptation of machine learning models and global improvement of the machine learning models while ensuring data security and privacy. An artificial reality headset uses smart sensors and a system on chip (SoC). The smart sensor extracts features from sensor data, encrypts them and transmits the encrypted features to the SoC of the headset. The SoC provides the encrypted features to a server, for example, a server in a cloud platform. The server updates a base model with collected features and sends the updated base model parameters to headsets. The smart sensors of a headset receiving the updated model parameters, balance the model parameters with locally stored feature data, for example, user calibration features to perform user customization of the machine learning model.

Federated on-sensor learning for local device adaptation and global improvement of machine learning model

A console and headset system locally trains machine learning models to perform customized online learning tasks. To customize the online learning models for specific users of the system without using outside resources, the system trains the models to compare a target frame to stored calibration frames, rather than directly inferring information about a target frame. During deployment, an embedding is generated for the target frame. A sample embedding that is closest to the target embedding is selected from a group of embeddings of calibration frames. The information about the selected embedding and target embedding and ground truths for the calibration frame are provided as inputs to one of the trained models. The model predicts a difference between the target frame and the calibration frame, which can be used to determine information about the target frame.

Personalized online learning for artificial reality applications

An example apparatus having a heterogenous memory system includes a first sensor layer, of a plurality of stacked sensor layers, including an array of pixels; and one or more semiconductor layers of the plurality of stacked sensor layers located beneath the first sensor layer, the one or more semiconductor layers configured to process pixel data output by the array of pixels, the one or more semiconductor layers including a first memory to store most significant bits ('MSBs') of data involved in the processing of the pixel data; a second memory to store least significant bits (“LSBs”) of the data; and wherein the first memory has a lower bit error rate (“BER”) than the second memory.

Heterogeneous memory system

In one example, an apparatus comprises: a first sensor layer, including an array of pixel cells configured to generate pixel data; and one or more semiconductor layers located beneath the first sensor layer with the one or more semiconductor layers being electrically connected to the first sensor layer via interconnects. The one or more semiconductor layers comprises on-chip compute circuits configured to receive the pixel data via the interconnects and process the pixel data, the on-chip compute circuits comprising: a machine learning (ML) model accelerator configured to implement a convolutional neural network (CNN) model to process the pixel data; a first memory to store coefficients of the CNN model and instruction codes; a second memory to store the pixel data of a frame; and a controller configured to execute the codes to control operations of the ML model accelerator, the first memory, and the second memory.

Growing community of inventors

Belmont, CA, United States of America

Barbara De Salvo