Neural Network Signal Propagation via Discrete Time Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for processing input signals through neural networks in machine learning systems are inefficient due to sequential processing, requiring repeated initialization and lacking effective parallelization, which hampers computational speed and memory efficiency.
Innovation Solution
A method that allows for parallel propagation of signals through layers of a neural network at discrete time increments, enabling efficient computation by using pre-defined time pulses and recurrent layers, and associating different computation steps with different processing units, allowing for partial or complete parallelization of operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If sequential processing is used to propagate signals through neural network layers, then memory efficiency is improved, but computational speed deteriorates
Solution Approach 1:
The patent segments the neural network computation into discrete time increments, where each time increment processes a specific set of layers. This segmentation allows parallel processing of different layers at the same time increment while maintaining organized memory access patterns, thus improving computational speed without sacrificing memory efficiency.
Solution Approach 2:
The patent introduces dynamic time increments that allow the system to adaptively control the propagation of signals through layers. By using dynamic time management, the system can parallelize computations across multiple layers simultaneously while maintaining efficient memory utilization through controlled signal propagation timing.
2Measurement precision
If repeated initialization is performed for each input signal, then processing accuracy is improved, but processing time deteriorates
Solution Approach 1:
The patent performs initialization only once before processing the sequence of input signals, rather than repeating it for each signal. This preliminary action maintains processing accuracy by ensuring proper initialization while dramatically reducing processing time by eliminating redundant initialization operations.
Solution Approach 2:
The patent maintains continuous processing of input signals through the neural network after initial initialization, allowing the system to process sequences of signals without interruption or repeated initialization. This continuity preserves accuracy while minimizing time loss by keeping the network in an active processing state.
3Productivity
If parallel processing is implemented across layers, then computational speed is improved, but memory efficiency deteriorates
Solution Approach 1:
The patent segments parallel processing into discrete time increments, where each increment handles specific layers. This segmentation enables parallel computation across multiple layers while maintaining efficient memory usage by processing layers in organized batches rather than all simultaneously, thus improving speed without proportionally increasing memory requirements.
4Adaptability or versatility
If computation time varies for different layers, then processing flexibility is improved, but deterministic results deteriorate
Solution Approach 1:
The patent introduces periodic time increments that structure the parallel processing of layers in a regular, predictable pattern. This periodic action maintains processing flexibility by allowing different layers to be processed at different rates while ensuring deterministic results through the regular timing structure that synchronizes signal propagation across the network.
Data Source
AI summary
A method for efficiently ascertaining output signals of a sequence of output signals with the aid of a sequence of layers of a machine learning system, in particular a neural network, from a sequence of input signals. The neural network is supplied in succession with the input signals of the sequence of input signals in a sequence of discrete time increments. At the discrete time increments, signals present in the network are in each case further propagated through a layer of the sequence of layers.