Spiking Neural Network Sub-Assembly Sizing for Data Recall
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Spiking neural networks face challenges in effectively memorizing and recalling data signals due to limitations in learning and reproducing spatio-temporal spike patterns, leading to inefficiencies in data storage and retrieval.
Innovation Solution
A method is introduced that involves using a machine learning model to classify data signals based on metadata, dynamically configuring sub-assemblies of neurons in the spiking neural network to prioritize the storage of more relevant signals, allowing for larger sub-assemblies to store higher relevance data, and employing spike-frequency encoding to enhance the fidelity of data storage and recall.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a fixed-size sub-assembly of neurons is used to memorize data signals, then the network structure is simple, but the reliability of data recall deteriorates because all data signals are treated equally regardless of their importance
Solution Approach 1:
The patent applies dynamics by making the sub-assembly size variable rather than fixed. The classifier dynamically determines the relevance class of each data signal, and the controller dynamically configures the sub-assembly size based on this relevance class. More relevant data signals are stored in larger sub-assemblies, improving recall reliability, while less relevant signals use smaller sub-assemblies, managing complexity.
Solution Approach 2:
The patent applies local quality by differentiating the storage quality for different data signals based on their relevance. Instead of uniform treatment, each data signal receives a locally optimized sub-assembly size proportional to its importance. This ensures high-reliability recall for critical data while maintaining efficiency for less important data.
2Manufacturing precision
If larger sub-assemblies are used to store more relevant data signals, then the fidelity of data storage is improved, but the use of neural resources deteriorates due to increased neuron allocation
Solution Approach 1:
The patent applies parameter changes by varying the sub-assembly size parameter based on the relevance class of each data signal. The classifier outputs a relevance class that serves as a parameter to control the number of neurons allocated. This ensures high-fidelity storage for relevant data while optimizing neural resource utilization across the entire network.
Solution Approach 2:
The patent applies discarding and recovering by allowing neurons to be dynamically allocated and deallocated based on data relevance. Less relevant data signals use smaller sub-assemblies, effectively 'discarding' the use of excess neurons that would otherwise be wasted. These neurons can then be recovered and reused for other data storage tasks, optimizing overall resource efficiency.
3Reliability
If uniform sub-assembly size is used for all data signals, then the neural resource allocation is efficient, but the reliability of recalling relevant information deteriorates
Solution Approach 1:
The patent applies local quality by providing differentiated sub-assembly sizes tailored to the specific needs of each data signal's relevance. Critical information receives enhanced storage resources (larger sub-assemblies) to ensure reliable recall, while non-critical information uses standard resources. This local optimization improves overall system reliability without uniform resource expenditure.
Solution Approach 2:
The patent applies parameter changes by using the relevance class parameter to control sub-assembly configuration. The classifier generates a relevance parameter that dynamically adjusts the storage allocation. This parameter-driven approach ensures that neural resources are allocated efficiently according to data importance, improving both reliability and productivity.
4Adaptability or versatility
If the spiking neural network dynamically reconfigures sub-assemblies based on relevance classes, then the adaptability of the network is improved, but the complexity of control mechanisms deteriorates
Solution Approach 1:
The patent applies segmentation by dividing the control mechanism into two distinct functional modules: a classifier that assesses data relevance and a controller that configures sub-assemblies based on classifier output. This segmentation simplifies the overall control complexity by separating the decision-making (relevance assessment) from the execution (sub-assembly configuration), while maintaining high adaptability.
Data Source
AI summary
Methods and apparatus are provided for memorizing data signals in a spiking neural network. For each data signal, such a method includes supplying metadata relating to the data signal to a machine learning model trained to generate an output signal, indicating a relevance class for a data signal, from input metadata for that data signal. The method includes iteratively supplying the data signal to a sub-assembly of neurons, interconnected via synaptic weights, of a spiking neural network and training the synaptic weights to memorize the data signal in the sub-assembly. The method further comprises assigning neurons of the network to the sub-assembly in dependence on the output signal of the model such that more relevant data signals are memorized by larger sub-assemblies. The data signal memorized by a sub-assembly can be subsequently recalled by activating neurons of that sub-assembly.