Term #23
Attribute (long)
Claim Term
Attribute (long)
Origin matter
Placeholder App
Reference Case 1
FSP1763
Date Added
1/1/20
Created
1/1/20, 12:00 AM
Modified
8/30/21, 3:47 PM
Full Desc
As used herein, "attribute" refers to any property, trait, aspect, quality, data value, setting, or feature of an object or thing. In embodiments of the claimed solution, attribute refers to properties of an object detector and may include, but is not limited to, an accuracy level for the object detector, a latency for the object detector between receiving input(s) and providing an output (e.g., an inference result, an object detection prediction), a measure of an amount of memory resources the object detector uses, a measure of a resolution level for an image or frame provided as input to the object detector, a measure of an amount of processor resources the object detector uses, a measure of the number of computations the object detector performs per unit of time, such as seconds, and the like.
Where the object detector comprises a neural network, the attribute(s) of the object detector may include, but are not limited to, a type of neural network, a number of layers of the neural network, a number of nodes of the neural network, a number and/or type of interconnection between nodes of the neural network, a number of parameters used in the neural network, a number of floating point operations per second (FLOPS) for the neural network, and the like.
Where the object detector comprises a neural network, object detectors may be compared based on attributes for each object detector. In certain embodiments, object detectors in the form of neural networks may be compared, at a high level, using a rough comparison reference to size or weight. Generally, these size or weight comparisons of neural networks may be used to compare the neural networks based on a collection of attributes that relate to tradeoffs between one or more performance metrics and one or more operational constraints.
For example, an object detector/neural network may be described as heavy, heavyweight, large, thick, or fat and have the attributes of having a relatively high number of nodes, high number of layers, high FLOPS, high memory usage, and/or high computational latency, in exchange for higher accuracy of object detection.
In contrast and by comparison, another object detector/neural network may be described as light, lightweight, small, thin, or lean and have the attributes of having a relatively small/low number of nodes, small/low number of layers, small/low FLOPS, small/low memory usage, and/or small/low computational latency, in exchange for lower accuracy of object detection.
Where the object detector comprises a neural network, and a convolutional neural network in particular, the attribute(s) may also be referred to as hyperparameters and may include aspects such as a number of total layers to use in the neural network, a number of convolution layers, filter sizes, values for strides at each layer, and/or the like.
Applications using this term
| Matter | Usage | Notes | Actions |
|---|---|---|---|
| Placeholder App | Defined | — | App terms |