Real-Time Temporal Convolution Network Algorithm
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for real-time execution of Temporal Convolution Networks (TCNs) on devices with limited compute bandwidth and memory are challenging due to the need for redundant convolution operations.
Innovation Solution
The proposed algorithm, RT-TCN, reuses the output of prior convolution operations in all layers of the network, minimizing computational requirements and memory footprint by performing convolution operations only when all necessary inputs are available and retaining necessary input and intermediate values in memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If redundant convolution operations are performed to compute latest output in real-time, then the TCN can be executed continually as new data becomes available, but computational resources and memory bandwidth are consumed excessively
Solution Approach 1:
The patent performs preliminary convolution operations and stores their outputs in memory before they are needed for subsequent predictions. By pre-computing and retaining intermediate convolution results, the system avoids redundant computations when new data arrives, enabling efficient real-time evaluation without excessive computational resource consumption
Solution Approach 2:
The patent creates a simplified model that copies only the necessary intermediate convolution outputs from the full TCN and stores them in memory. This selective copying approach allows the system to retain useful computational results while minimizing memory usage and avoiding the need to re-perform all convolution operations from scratch
2Productivity
If all inputs comprising the receptive field are stored in memory for real-time execution, then the TCN can compute outputs efficiently, but memory requirements become excessive
Solution Approach 1:
The patent extracts only the necessary intermediate convolution outputs from the full set of possible inputs and stores just these critical intermediates in memory. By taking out only the essential computational results rather than storing all possible inputs, the system reduces memory requirements while maintaining the ability to compute outputs efficiently
Solution Approach 2:
The patent changes the parameter of what is stored in memory from raw input data to processed intermediate convolution outputs. This parameter transformation allows the system to store compressed representations of computational state that require less memory while preserving the information needed for real-time computation
Data Source
AI summary
A real-time temporal convolution network (RT-TCN) algorithm reuses the output of prior convolution operations in all layers of the network to minimize the computational requirements and memory footprint of a TCN during real-time evaluation. Further, a TCN trained via the fixed-window view, where the TCN is trained using fixed time splices of the input time series, can be executed in real-time continually using RT-TCN.