Prior art — US Patent 11775831

To identify the most relevant prior art for US patent 11775831, I have examined the patent citations listed within the provided authoritative patent text. Only citations with a publication date prior to the priority date of US11775831 (September 26, 2016) are considered for potential anticipation under 35 U.S.C. § 102.

Due to limitations in retrieving the full abstracts of the cited patents, the "brief description" for each reference below is derived from its title as presented in the US11775831 patent document. A definitive assessment of anticipation would typically require a review of the full patent text, including the claims and detailed description of each cited prior art reference.

Here are the patent citations from US11775831 that qualify as prior art based on their publication date:

Relevant Prior Art for US11775831

US5822452A
- Full Citation: US5822452A - System and method for narrow channel compression
- Publication Date: 1998-10-13
- Assignee: 3Dfx Interactive, Inc.
- Brief Description: This patent describes a system and method for narrow channel compression.
- Potential Anticipation (35 U.S.C. § 102): Based on the title, this patent appears to relate to general data compression techniques. The core innovation of US11775831 lies in its specific application of cascaded computing within Convolutional Neural Networks (CNNs) by initially using Most Significant Bits (MSBs) to identify a maximum value and then performing full-precision computation only on the data set exhibiting that maximum, particularly within convolution and pooling operations. A general compression method is unlikely to disclose these specific steps in the context of CNNs and cascaded precision. Therefore, it is unlikely to anticipate Claim 1 or its dependent claims.
US20150032449A1
- Full Citation: US20150032449A1 - Method and Apparatus for Using Convolutional Neural Networks in Speech Recognition
- Publication Date: 2015-01-29
- Assignee: Nuance Communications, Inc.
- Brief Description: This patent discloses a method and apparatus for using Convolutional Neural Networks in speech recognition.
- Potential Anticipation (35 U.S.C. § 102): While this patent involves CNNs, its focus is on their application in speech recognition. The title does not indicate any disclosure of the cascaded computing methodology of US11775831, which involves initial low-precision computation using MSBs, determination of a maximum, and subsequent selective full-precision computation for efficiency in CNN layers. Thus, it is unlikely to anticipate Claim 1.
US20150255062A1
- Full Citation: US20150255062A1 - System and method for applying a convolutional neural network to speech recognition
- Publication Date: 2015-09-10
- Assignee: Gerald Bradley PENN
- Brief Description: This patent describes a system and method for applying a convolutional neural network to speech recognition.
- Potential Anticipation (35 U.S.C. § 102): Similar to US20150032449A1, this reference details the use of CNNs in speech recognition. The title does not suggest the specific computational efficiency techniques of US11775831, which are centered around cascaded precision and selective computation based on maximum values in pooling layers. Therefore, it is unlikely to anticipate Claim 1.
US20150339571A1
- Full Citation: US20150339571A1 - System and method for parallelizing convolutional neural networks
- Publication Date: 2015-11-26
- Assignee: Google Inc.
- Brief Description: This patent describes a system and method for parallelizing convolutional neural networks.
- Potential Anticipation (35 U.S.C. § 102): This patent addresses computational efficiency by parallelizing CNN operations. While both patents aim for efficiency, the method in US11775831 specifically reduces the total amount of computation by avoiding redundant calculations through a cascaded precision approach using MSBs and selective full-precision. Parallelization, while beneficial, does not inherently teach the specific steps of performing initial low-precision calculations, identifying a maximum, and then selectively applying full-precision computation. Thus, it is unlikely to anticipate Claim 1.
WO2016033506A1
- Full Citation: WO2016033506A1 - Processing images using deep neural networks
- Publication Date: 2016-03-03
- Assignee: Google Inc.
- Brief Description: This patent broadly relates to processing images using deep neural networks.
- Potential Anticipation (35 U.S.C. § 102): This broad title indicates a patent related to a general application of deep neural networks, which can include CNNs, for image processing. However, it does not explicitly suggest the specific cascaded computing approach described in US11775831, which focuses on iteratively increasing precision (starting with MSBs) and performing full-precision computation only on data sets that exhibit maximum values in pooling operations to reduce overall computation. Therefore, it is unlikely to anticipate Claim 1.
US9436895B1
- Full Citation: US9436895B1 - Method for determining similarity of objects represented in images
- Publication Date: 2016-09-06
- Assignee: Mitsubishi Electric Research Laboratories, Inc.
- Brief Description: This patent describes a method for determining the similarity of objects represented in images.
- Potential Anticipation (35 U.S.C. § 102): This patent's subject matter is image object similarity. While CNNs could be used for this purpose, the title does not suggest the specific computational efficiency methodology employed by US11775831, which involves cascaded precision based on MSBs and selective full-precision computation to optimize pooling operations. Hence, it is unlikely to anticipate Claim 1.

Based on the available information (titles of cited patents), none of the identified prior art references explicitly disclose all the elements of independent Claim 1 of US11775831, particularly the specific combination of performing an initial computation based on a first set of MSBs, examining for a maximum, and then conditionally performing a full-precision computation only for the data set exhibiting that maximum, and propagating that full precision value to a subsequent layer within a CNN. The cited references generally address broader aspects of CNNs, their applications, or different efficiency mechanisms.