LSTM State Control Device for Preventing Unintentional Reset
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks, such as LSTM models, face issues with state variable reset due to frequent inputs, leading to insufficient storage duration, where the state value is unintentionally reset within a few seconds, even when 120 inputs are performed per second.
Innovation Solution
A state control device that includes an input data acquiring section, a processing executing section, and a state control section to control the update of the neural network's state by determining whether to input data to the network or return it to a previous state, based on input and output data, using an input determination model for learning and controlling the update process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If frequent input is performed to the neural network, then processing speed is improved, but the state is reset unintentionally and storage duration is reduced
Solution Approach 1:
An input determination model is introduced as an intermediary between the input data and the neural network. This mediator evaluates whether each input should trigger a state update, allowing frequent inputs to be processed without causing unwanted state resets. The intermediary filters inputs based on learned criteria, maintaining both high processing speed and appropriate state storage duration.
Solution Approach 2:
The system dynamically changes the parameter of state update frequency by using the input determination model to selectively allow or block state updates. Instead of updating state on every input (fixed parameter), the system adapts the update frequency based on the content and context of each input, optimizing both processing speed and state retention.
2Measurement precision
If the neural network updates state on every input, then processing accuracy is improved, but computational overhead increases
Solution Approach 1:
Instead of updating the state on every single input (excessive action), the system applies partial updates only when the input determination model deems it necessary. This selective approach maintains processing accuracy for critical inputs while reducing unnecessary computational overhead from redundant state updates.
Solution Approach 2:
The input determination model learns from the neural network's outputs and automatically determines when state updates are needed, making the system self-regulating. This self-service mechanism reduces the need for external control logic and minimizes computational overhead while maintaining accuracy.
3Duration of action of stationary object
If the state is restricted from updating, then state storage duration is improved, but processing adaptability deteriorates
Solution Approach 1:
The system transitions from a static state update mechanism (always update or never update) to a dynamic one where the update decision adapts based on the input characteristics. The input determination model learns optimal update conditions, allowing the state storage duration to be extended while maintaining processing adaptability through context-aware update selection.
Solution Approach 2:
The input determination model uses feedback from the neural network's outputs to make informed decisions about state updates. By monitoring the relationship between inputs and outputs, the system determines when state updates will be beneficial, thus extending storage duration without sacrificing adaptability to changing conditions.
Data Source
AI summary
A target input data acquiring section acquires target input data. A processing executing section executes processing that uses output data which is an output from an LSTM model into which the target input data is input. The loop processing including the acquisition of the target input data and the processing execution is repeatedly executed. A state control section controls whether or not to restrain the update of the states associated with the LSTM model, on the basis of at least one of the input data and the output data.