Obviousness

Obviousness Analysis of US Patent 11775831 under 35 U.S.C. § 103

This analysis identifies combinations of prior art references that would render the claims of US Patent 11775831 obvious to a person having ordinary skill in the art (PHOSITA) as of the priority date of September 26, 2016. The core of the invention lies in using cascaded, low-precision computations (starting with most significant bits, MSBs) in Convolutional Neural Networks (CNNs) to efficiently identify a maximum value in a pooling operation, and then performing full-precision computation only on the data set that exhibited this maximum, thereby reducing overall computational load without affecting accuracy.

Independent Claim 1 of US11775831 Recites:

1. One or more non-transitory computer-readable storage media storing instructions which, when executed by at least one processor, cause the at least one processor to perform operations comprising:
   • in one or more layers of a convolutional neural network (CNN), performing a first iteration that includes computing a value based on a first set of most significant bits (MSBs) for each of a plurality of data sets;
   • examining a first set of values computed for the plurality of data sets in the first iteration to determine whether a maximum value is present among the first set of values;
   • responsive to identifying the maximum value, performing a full precision computation of the value for a data set, of the plurality of data sets, that exhibited the maximum value; and
   • propagating the full precision computation of the value to a subsequent layer of the CNN.

Identified Prior Art References and Their Relevance (Pre-September 26, 2016):

The following prior art references, listed in the "Citations" section of US11775831, are relevant due to their publication or priority dates being before September 26, 2016:

• [A] WO2016033506A1 (Google Inc., published March 3, 2016): Titled "Processing images using deep neural networks," this reference discloses the use of deep neural networks, which a PHOSITA would understand to encompass Convolutional Neural Networks (CNNs) with their characteristic layers, including convolution and pooling operations, for tasks like image processing. In CNNs, data is processed through layers, and the output of one layer is propagated to the next. Pooling layers, particularly max-pooling, are standard for identifying maximum values among a plurality of data sets to reduce spatial dimensions.
• [B] US20170300815A1 (Arizona Board of Regents On Behalf Of Arizona State University, priority date April 13, 2016): Titled "Static and dynamic precision adaptation for hardware learning and classification," this reference explicitly teaches methods for reducing computational cost and improving efficiency in neural networks through "precision adaptation." It describes dividing input features into "groups of precision values" and performing "cascaded operations." A PHOSITA would infer that "groups of precision values" for initial, coarse computations would logically refer to most significant bits (MSBs), and "cascaded operations" imply an iterative refinement process.

Obviousness Combination: [A] WO2016033506A1 in view of [B] US20170300815A1

A PHOSITA would have found Independent Claim 1 of US11775831 obvious by combining the teachings of WO2016033506A1 and US20170300815A1.

Motivation for Combination:

The problem addressed by US11775831—that CNNs are "computationally and memory intensive" and "may involve many redundant operations" (US11775831, Detailed Description)—was a well-known challenge in the field of neural networks at the priority date of this patent. The patent explicitly states that a "largest redundancies in conventional CNNs is that a large amount of data is thrown away at each pooling layer, because only the maximum value is conveyed to the next layer" (US11775831, Detailed Description).

WO2016033506A1 teaches the use of CNNs, which involve computationally intensive convolution and pooling operations. US20170300815A1, from the same original assignee as US11775831, directly addresses the need to reduce computational cost and improve efficiency in neural networks by employing "precision adaptation" and "cascaded operations" with varying "groups of precision values."

Given the known computational burden of CNNs (as taught by [A]) and the explicit teaching of reducing computational cost through precision adaptation and cascaded operations (as taught by [B]), a PHOSITA would have been highly motivated to combine these teachings. The specific motivation would be to apply the precision adaptation techniques of [B] to the convolution and pooling operations within the CNNs of [A] to overcome the known redundancy issue where most computed values are discarded after pooling. It would be an obvious design choice to perform initial computations with lower precision (e.g., MSBs) to quickly identify the likely maximum before investing resources in full-precision computation.

Mapping Claim 1 Elements to the Combination:

• "in one or more layers of a convolutional neural network (CNN)": WO2016033506A1 explicitly discloses "deep neural networks" for "processing images," which are understood by a PHOSITA to include CNNs with multiple layers.
• "performing a first iteration that includes computing a value based on a first set of most significant bits (MSBs) for each of a plurality of data sets": US20170300815A1 teaches "precision adaptation" in neural networks by dividing "input features... into a group of precision values" and performing "cascaded operations" to reduce computational cost. A PHOSITA, seeking to perform an initial, coarse computation for efficiency, would naturally select the most significant bits (MSBs) as the "first set of bits" for this initial approximate calculation of values for a plurality of data sets within a CNN layer as taught by [A].
• "examining a first set of values computed for the plurality of data sets in the first iteration to determine whether a maximum value is present among the first set of values": This step involves a pooling operation, particularly max-pooling, which is a standard component of CNNs as taught by WO2016033506A1. A PHOSITA would apply this standard pooling operation to the results of the MSB-based computations performed as described above.
• "responsive to identifying the maximum value, performing a full precision computation of the value for a data set, of the plurality of data sets, that exhibited the maximum value": US20170300815A1's goal of "precision adaptation" to "reduce computational cost" inherently teaches performing higher-precision computation only when necessary. If the initial low-precision (MSB-based) calculation clearly identifies a maximum (as occurs in pooling), it would be an obvious implementation of efficiency to perform the more resource-intensive "full precision computation" exclusively on that single data set that exhibited the maximum, and not on the other data sets that are destined to be discarded by the pooling layer. This directly addresses the redundancy described in US11775831 where 75-89% of data is typically thrown away.
• "propagating the full precision computation of the value to a subsequent layer of the CNN.": This is a fundamental operation in any multi-layer neural network, including the CNNs disclosed by WO2016033506A1.

Addressing Dependent Claims 2, 3, and 11:

• Claims 2 and 3 (Second iteration with larger set of MSBs): US20170300815A1's teaching of "cascaded operations" and "precision adaptation" using "groups of precision values" directly anticipates a scenario where an initial low-precision computation is insufficient (e.g., if multiple values are too similar to confidently determine a maximum). In such a case, a PHOSITA would find it obvious to proceed with a subsequent iteration using a larger "group of precision values" (i.e., more MSBs) to refine the computation until a clear maximum is identified, consistent with the goal of "precision adaptation." This iterative refinement is a known technique for approximate computing.
• Claim 11 (Full precision computation on less data): This is inherently covered by the motivation described above. The explicit goal of US20170300815A1 to reduce computational cost through precision adaptation, when applied to a CNN with pooling (WO2016033506A1), would lead a PHOSITA to perform full precision computation only on the data selected by the pooling layer (the maximum), which by definition is a subset ("less data") of the initial plurality of data sets.

Conclusion:

The combination of WO2016033506A1, which discloses Convolutional Neural Networks with pooling operations, and US20170300815A1, which teaches precision adaptation and cascaded operations in neural networks to reduce computational cost, would render the subject matter of Independent Claim 1 (and its dependent claims) of US11775831 obvious to a PHOSITA. The pervasive problem of computational inefficiency and redundancy in CNNs, particularly in pooling layers, would provide ample motivation to combine these references to achieve a more efficient CNN architecture by selectively applying full-precision computation only to the data that truly matters for subsequent layers.