Neuron Calculator for Wireless Transceivers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing complexity and memory/power demands of larger neural networks in applications like image and speech recognition, particularly in 5G wireless communication systems, lead to inefficiencies due to exponential growth in the number of weights and neurons, which conventional pruning or compression methods often fail to address without introducing additional complexity.
Innovation Solution
The system calculates ordered sets based on input signals to form neural networks that utilize these signals for memory allocation, reducing the number of elements and facilitating more efficient memory use, while also increasing processing speed by leveraging known input signals, using a neuron calculator to determine connection weights through algorithms like least-mean squares and gradient descent.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of neurons and weights in neural networks is increased to improve learning capability, then the accuracy and performance of image and speech recognition are improved, but the memory and power consumption of devices are overwhelmed
Solution Approach 1:
The patent extracts and removes redundant weights and neurons from the neural network through pruning techniques. By identifying and eliminating unnecessary connections between neurons, the system reduces the total number of weights while preserving the essential learning capabilities, thereby reducing memory consumption without significantly compromising recognition accuracy
Solution Approach 2:
The patent changes the parameter of weight precision by reducing it from high-precision representations to lower-precision formats. This parameter change allows the neural network to maintain adequate performance while occupying significantly less memory space, resolving the contradiction between accuracy and memory consumption
2Measurement precision
If the number of neurons and weights is increased to improve learning capability, then the performance of neural networks is improved, but the power consumption of devices increases
Solution Approach 1:
By pruning redundant neurons and weights, the patent reduces the total computational operations required during forward and backward propagation. Fewer operations mean less energy consumption, allowing the system to maintain high recognition accuracy with reduced power usage
Solution Approach 2:
The reduction in weight precision parameters decreases the computational complexity of arithmetic operations. Lower-precision calculations require fewer computational cycles and less energy, thereby reducing power consumption while maintaining adequate recognition performance
3Quantity of substance
If conventional pruning or compression methods are applied to reduce the number of weights, then memory consumption is reduced, but additional complexity is introduced to the neural network
Solution Approach 1:
The patent merges the pruning process with the standard training workflow by integrating weight importance evaluation into the existing backpropagation algorithm. This integration allows the system to identify and remove redundant weights using the same computational framework already in place, avoiding the introduction of separate complex pruning modules
Solution Approach 2:
The neural network performs self-pruning by automatically identifying redundant weights through its own training process. The backpropagation algorithm naturally reveals which weights have minimal impact on the loss function, allowing the network to self-optimize its structure without external intervention or additional complexity
Data Source
AI summary
Examples described herein include systems and methods, including wireless devices and systems with neuron calculators that may perform one or more functionalities of a wireless transceiver. The neuron calculator calculates output signals that may be implemented, for example, using accumulation units that sum the multiplicative processing results of ordered sets from ordered neurons with connection weights for each connection between an ordered neuron and outputs of the neuron calculator. The ordered sets may be a combination of some input signals, with the number of signals determined by an order of the neuron. Accordingly, a kth-order neuron may include an ordered set comprising product values of k input signals, where the input signals are selected from a set of k-combinations with repetition. As an example in a wireless transceiver, the neuron calculator may perform channel estimation as a channel estimation processing component of the receiver portion of a wireless transceiver.