Neuromorphic Neural Network Device Time-Delayed Convolutions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network devices face challenges in efficiently processing large amounts of input data in real-time while maintaining low power consumption, particularly in analyzing and extracting information using convolutional neural networks (CNNs).
Innovation Solution
A neural network device incorporating neuromorphic hardware with variable resistive devices that switch resistance states based on applied voltage or current, allowing for efficient CNN mapping by performing operations between input data and kernels, with the ability to handle delayed input data streams and generate output data using results from multiple kernel operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional neural network devices process large amounts of input data in real-time, then processing speed and productivity are improved, but power consumption increases
Solution Approach 1:
The patent segments the input data stream into multiple parallel streams with different time delays. Each stream is processed independently through the same kernel, allowing the system to handle multiple input frames simultaneously. This segmentation enables efficient utilization of hardware resources and reduces the need for sequential processing, thereby improving productivity while maintaining low power consumption through parallel operations.
Solution Approach 2:
The patent implements periodic action by creating data streams with specific time delays (e.g., first stream with delay T1, second stream with delay T2). These periodic delayed streams are processed in an interleaved manner, allowing the neural network to maintain a steady processing rhythm. This periodic structure enables efficient pipeline utilization and keeps the hardware operating at optimal efficiency points, improving productivity without proportionally increasing power consumption.
2Adaptability or versatility
If the neural network device uses multiple input streams with different delay times, then adaptability is improved, but device complexity increases
Solution Approach 1:
The patent applies universality by designing a single processing core that handles multiple input streams with different delay characteristics. The same kernel and processing logic are reused across all streams, making the device versatile in handling various temporal patterns. This multi-functional approach allows the system to adapt to different input scenarios without requiring separate dedicated hardware for each stream, thereby limiting the increase in device complexity.
Solution Approach 2:
Instead of creating entirely separate processing paths for each delayed stream, the patent uses copying by replicating the input data with different time delays and routing them through shared processing resources. The kernel weights and processing logic are copied/reused across multiple streams rather than duplicating entire processing pipelines. This approach enables adaptability to handle multiple delay patterns while keeping the overall device complexity manageable through resource sharing.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach enables efficient real-time processing of large input data sets with reduced power consumption, effectively handling image data streams with different delay times and improving the performance of CNN operations in neural network devices.
Implementation Method 1
each of the cores comprise variable resistive devices configured to switch between different resistance states based on a voltage or a current applied to both ends of the variable resistive device
Implementation Method 2
A conductivity of the variable resistive devices may change, in response to a voltage greater than a threshold being applied to the both ends of the variable resistive device
Data Source
AI summary
A method of performing operations on a plurality of inputs and a same kernel using a delay time by using a same processor, and a neural network device thereof are provided, the neural network device includes input data including a first input and a second input, and a processor configured to obtain a first result by performing operations between the first input and a plurality of kernels, to obtain a second result by performing operations between the second input, which is received at a time delayed by a first interval from a time when the first input is received, and the plurality of kernels, and to obtain output data using the first result and the second result. The neural network device may include neuromorphic hardware and may perform convolutional neural network (CNN) mapping.