Neural Network Acceleration With Discretized Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks require significant computational resources and memory, limiting their use on resource-constrained devices due to high precision weights, which complicates real-time performance and power consumption.
Innovation Solution
Implementing discretized weights with reduced precision (binary, ternary, or quaternary values) and using a control unit to coordinate datapath operations for convolution and fully connected layers, allowing for efficient computation and reduced memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high precision weights are used in neural networks, then computational accuracy is improved, but memory requirements and power consumption increase
Solution Approach 1:
The patent applies parameter changes by transitioning from high precision (e.g., 32-bit floating point) weights to discretized weights with reduced precision (binary, ternary, or quaternary values). This changes the representation parameter of weights while maintaining functional accuracy through carefully designed discretization schemes that preserve the essential computational behavior of the neural network.
Solution Approach 2:
The patent creates a simplified copy of the weight data structure by replacing full-precision floating point numbers with discretized representations. This copy maintains the functional essence of the weights while using significantly less memory space, enabling deployment on resource-constrained devices.
2Measurement precision
If high precision weights are used in neural networks, then computational accuracy is improved, but power consumption increases
Solution Approach 1:
The patent changes the precision parameter of weight representations from high precision floating point to low precision discretized values. This parameter change reduces the computational workload and data movement requirements, thereby lowering power consumption while maintaining acceptable computational accuracy for the application.
3Quantity of substance
If discretized weights with reduced precision are used, then memory requirements and power consumption are reduced, but computational accuracy may deteriorate
Solution Approach 1:
The patent carefully controls the parameter changes in weight discretization by selecting appropriate discretization levels (binary, ternary, quaternary) and strategies that minimize accuracy loss. The discretization is designed to preserve the essential computational characteristics needed for the specific neural network application.
Solution Approach 2:
The patent creates a simplified copy of the weight data that maintains functional equivalence to the original high-precision weights. Through careful copying and transformation of weight values into discretized forms, the patent preserves computational accuracy while reducing memory requirements.
4Productivity
If discretized weights are used, then real-time performance on resource-constrained devices is enabled, but device complexity increases
Solution Approach 1:
The patent changes the weight representation parameter to enable efficient computation on resource-constrained devices. The discretized weight format allows for simpler and faster arithmetic operations that can be implemented with basic hardware units, achieving real-time performance without requiring complex specialized hardware.
Data Source
AI summary
For neural network acceleration, a datapath can be configured to implement a convolution computation. A control unit can be configured to coordinate operations of the datapath to implement the convolution computation based on coded instructions representative of a neural network system. The control unit can be configured to command the datapath to convolve at least one input feature element of a set of input feature elements of at least one input feature map with at least one discretized weight of a set of discretized weights to compute an influence that the at least one input feature element of the set of input feature elements of the least one input feature map has on one or more output feature elements of at least one output feature map.