Dynamic Depth Field Number Format for Deep Learning Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models and accelerators face challenges with accuracy degradation when using smaller bitwidths in embedded/mobile environments, as existing fixed-point and floating-point number formats have limited representation ranges and bitwidths, leading to reduced accuracy in computations.
Innovation Solution
A computing apparatus that detects weight and activation depth fields within weight and activation values, generates operands, and performs operations to determine resultant depth fields, allowing for efficient data processing and output generation, even with smaller bitwidths, by using a novel number format that represents a wider range of values and more mantissa bits than traditional formats.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If fixed-point or floating-point number formats are used with smaller bitwidths, then energy consumption and data size are reduced, but computational accuracy degrades
Solution Approach 1:
The patent introduces a new number format dimension that combines depth field (exponent-like) and mantissa field (significand-like) in a structured arrangement. This dimensional reorganization allows the system to achieve high precision with reduced bitwidth by optimizing the allocation of bits across different dimensional components, thereby reducing energy consumption while maintaining accuracy.
Solution Approach 2:
The patent changes the parameter representation by using variable depth fields that can be detected and processed dynamically. By detecting the depth field (first target bit and subsequent bits) and adjusting the mantissa accordingly, the system adapts the precision parameters to match the actual computational needs, achieving high accuracy with smaller bitwidths and lower energy consumption.
2Quantity of substance
If smaller bitwidths are used for weight and activation values, then data size and communication traffic are reduced, but representation range and precision are limited
Solution Approach 1:
The patent segments the number representation into distinct components: a depth field (containing the first target bit and subsequent bits) and a mantissa field. This segmentation allows each component to be optimized independently - the depth field captures the magnitude information efficiently while the mantissa field provides the necessary precision, achieving both compact data size and wide representation range.
Solution Approach 2:
The system performs preliminary detection of the depth field (identifying the first target bit and processing subsequent bits) before performing the main computation. This preliminary action extracts and processes the magnitude information in advance, allowing the subsequent mantissa operations to focus only on the precision-critical parts, thereby maintaining representation range while reducing overall data size.
3Adaptability or versatility
If traditional number formats are used, then compatibility is maintained, but computational speed and efficiency are reduced
Solution Approach 1:
The patent implements a dynamic number format where the depth field and mantissa field can be detected and processed adaptively. The system dynamically identifies the first target bit and processes subsequent bits based on the actual data characteristics, enabling faster computation by avoiding fixed-format limitations while maintaining compatibility with existing systems through the flexible detection and processing mechanism.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A computing apparatus include a processing circuitry configured to detect a weight depth field, related to a range of a weight value of a plurality of weight values, within the weight value, and detect an activation depth field, related to a range of an activation value of a plurality of activation values, within the activation value; identify a first operand in the weight value, and identify a second operand in the activation value; and generate an output value having a resultant depth field determined based on the weight depth field and the activation depth field, by performing an operation based on the detected first and second operands.