Hybrid Memory Cell Units for Asynchronous RNN Data Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current recurrent neural networks (RNNs) face challenges in efficiently processing asynchronous sensory data from diverse sensors with different sampling rates, leading to high computational load and power consumption, and struggle to maintain accurate timing information in continuous time scenarios.
Innovation Solution
The implementation of a recurrent neural network (RNN) with hybrid memory cell units, including first and second memory cells, where the second memory cells are controlled by phase signals of an oscillatory frequency, allowing for selective updating based on phase signals and oscillation parameters, enabling efficient processing of asynchronous data with reduced computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional RNNs process asynchronous sensory data from diverse sensors with different sampling rates, then the system can handle multiple sensor inputs, but the computational load and power consumption increase significantly
Solution Approach 1:
The patent implements time gates that operate periodically with different oscillation frequencies to control when memory cells update their states. This periodic gating mechanism allows the network to process asynchronous sensor data at appropriate intervals rather than continuously, reducing computational load and power consumption while maintaining the ability to handle diverse sampling rates
Solution Approach 2:
The patent divides the memory system into hybrid memory cell units with separate first and second memory cells, each controlled by independent time gates with different oscillation frequencies. This segmentation allows different parts of the network to operate at different rates, matching the asynchronous nature of sensor inputs while avoiding unnecessary computations across the entire network
2Measurement precision
If conventional RNNs process data at high sampling rates, then timing information is captured more accurately, but computational overhead increases
Solution Approach 1:
The patent uses dynamic time gates with oscillation frequencies that can be adjusted based on the required timing precision and input characteristics. The time gates dynamically control when updates occur, providing high timing accuracy when needed while reducing computational overhead during periods where lower precision is acceptable
Solution Approach 2:
The patent changes the oscillation frequency parameters of time gates to match the sampling rates of different sensors and the required timing precision. By adjusting these parameters dynamically, the system maintains accurate timing information for critical operations while reducing computational overhead during less demanding periods
3Reliability
If time gates are fully opened for all memory cells, then all neurons are updated continuously, but computational load increases and convergence slows
Solution Approach 1:
The patent implements partial opening of time gates, where only a subset of memory cells are updated at each time step based on the oscillation phase and frequency. This partial action approach maintains reliability by ensuring all neurons are updated periodically while accelerating convergence by avoiding redundant updates to neurons that don't need immediate updates
Data Source
AI summary
A recurrent neural network including an input layer, a hidden layer, and an output layer, wherein the hidden layer includes hybrid memory cell units, each of the hybrid memory cell units including a first memory cells of a first type, the first memory cells being configured to remember a first cell state value fed back to each of gates to determine a degree to which each of the gates is open or closed, and configured to continue to update the first cell state value, and a second memory cells of a second type, each second memory cell of the second memory cells including a first time gate configured to control a second cell state value of the second memory cell based on phase signals of an oscillatory frequency, and a second time gate configured to control an output value of the second memory cell based on the phase signals, and each second memory cell of the second memory cells being configured to remember the second cell state value.