Microsoft Technology Licensing, LLC

Mohammed Shoaib

Jie Jerry Liu

Swagath Venkataramani

Jin Li

Nissanka Arachchige Bodhi Priyantha

Di Wang

Suman Kumar Nath

Dimitrios Lymberopoulos

Lei Zhang

Changhu Wang

Gerald Reuben DeJean

Yuxiao Hu

Michel Goraczko

Marcel Gavriliu

Scalable-effort machine learning may automatically and dynamically adjust the amount of computational effort applied to input data based on the complexity of the input data. This is in contrast to fixed-effort machine learning, which uses a one-size-fits-all approach to applying a single classifier algorithm to both simple data and complex data. Scalable-effort machine learning involves, among other things, classifiers that may be arranged as a series of multiple classifier stages having increasing complexity (and accuracy). A first classifier stage may involve relatively simple machine learning models able to classify data that is relatively simple. Subsequent classifier stages have increasingly complex machine learning models and are able to classify more complex data. Scalable-effort machine learning includes algorithms that can differentiate among data based on complexity of the data.

Scalable-effort classifiers for energy-efficient machine learning

Examples of the disclosure enable efficient processing of images. One or more features are extracted from a plurality of images. Based on the extracted features, the plurality of images are classified into a first set including a plurality of first images and a second set including a plurality of second images. One or more images of the plurality of first images are false positives. The plurality of first images and none of the plurality of second images are transmitted to a remote device. The remote device is configured to process one or more images including recognizing the extracted features, understanding the images, and/or generating one or more actionable items. Aspects of the disclosure facilitate conserving memory at a local device, reducing processor load or an amount of energy consumed at the local device, and/or reducing network bandwidth usage between the local device and the remote device.

Context-awareness through biased on-device image classifiers

Examples of the disclosure enable efficient processing of images. In some examples, one or more interest points are identified in an image. One or more features are extracted from the identified interest points using a filter module, a gradient module, a pool module, and/or a normalizer module. The extracted features are aggregated to generate one or more vectors. Based on the generated vectors, it is determined whether the extracted features satisfy a predetermined threshold. Based on the determination, the image is classified such that the image is configured to be processed based on the classification. Aspects of the disclosure facilitate conserving memory at a local device, reducing processor load or an amount of energy consumed at the local device, and/or reducing network bandwidth usage between the local device and the remote device.

Methods and systems for low-energy image classification

Disclosed herein are systems and methods for compressing data and for estimating sparsity of datasets to aid in compressing data. A device receives a plurality of samples of the sensor data from the sensor and determine a plurality of bits, in which each bit has a substantially equal probability of being determined as a 0 bit or of being determined as a 1 bit. The device estimates a sparsity value of the sensor data based at least in part on the sequence of bits. The device compresses the received samples of the sensor data based at least in part on the determined sparsity value to provide compressed data and transmits the compressed data via the transmitter to a receiver. Sparse data other than sensor data may also be compressed based at least in part on an estimated sparsity value.

Sparsity estimation for data transmission

Identifying products in a physical store shopping environment. The method includes, using a first detection method, identifying that a given product likely belongs to a given set of products. The method further includes, using one or more other detection methods, determining that the product is likely a specific product from the given set of products.

Multi-signal based shopping cart content recognition in brick-and-mortar retail stores

Systems, methods, and computer media for implementing convolutional neural networks efficiently in hardware are disclosed herein. A memory is configured to store a sparse, frequency domain representation of a convolutional weighting kernel. A time-domain-to-frequency-domain converter is configured to generate a frequency domain representation of an input image. A feature extractor is configured to access the memory and, by a processor, extract features based on the sparse, frequency domain representation of the convolutional weighting kernel and the frequency domain representation of the input image. The feature extractor includes convolutional layers and fully connected layers. A classifier is configured to determine, based on extracted features, whether the input image contains an object of interest. Various types of memory can be used to store different information, allowing information-dense data to be stored in faster (e.g., faster access time) memory and sparse data to be stored in slower memory.

Growing community of inventors

Redmond, WA, United States of America

Mohammed Shoaib