Neuromorphic Knowledge Graph Hardware for Real-Time Edge Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI systems for knowledge graph processing require extensive data extraction and offline training, making them unsuitable for real-time, online learning and inference in industrial settings like industrial automation systems, where data is sparse and dynamic.
Innovation Solution
The development of neuromorphic hardware and methods that enable online learning and inference using a probabilistic, sampling-based model for knowledge graphs, allowing for direct training on-edge without negative training examples, by switching between data-driven and model-driven learning modes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If classical statistical methods and offline training are used for knowledge graph processing, then training accuracy can be improved with sufficient data, but the system cannot perform real-time online learning and requires extensive data extraction and offline training infrastructure
Solution Approach 1:
The patent replaces classical statistical methods and offline training mechanisms with a neuromorphic hardware system that performs online learning through spiking neural networks. The mechanical process of data extraction, offline training, and model deployment is substituted with a biological-inspired system that learns continuously in real-time through spike-based neural computations, enabling both high accuracy and real-time productivity
Solution Approach 2:
The neuromorphic system enables self-service learning by processing and learning from data directly at the edge device without requiring external data extraction infrastructure or offline training pipelines. The system serves itself by continuously adapting its weights through online learning from incoming data streams, eliminating the need for separate data preparation and training phases
2Reliability
If extensive data extraction and offline training are performed, then comprehensive model training is achieved, but system complexity and latency increase, making it unsuitable for dynamic industrial settings
Solution Approach 1:
The patent segments the learning process into distributed spiking neural network units that operate independently and locally. Each neuron and synapse performs localized computations with simple update rules, avoiding the need for complex centralized training infrastructure. This segmentation enables reliable learning through simple, modular components that can be implemented in resource-constrained edge devices
Solution Approach 2:
The system employs periodic sampling of incoming data streams at discrete time intervals, processing data in batches rather than requiring continuous extensive data extraction. This periodic action allows the system to achieve reliable model training through regular, manageable data processing cycles, reducing system complexity compared to continuous comprehensive data extraction and offline training
3Stability of the object's composition
If offline training with pre-defined vocabularies is used, then structured knowledge representation is achieved, but the system cannot adapt to sparse and dynamic data in industrial automation systems
Solution Approach 1:
The patent implements dynamic adaptability through spiking neural networks that continuously update their weights and connections in response to incoming data. The system transitions from static pre-defined vocabularies to dynamic, evolving knowledge representations that adapt to sparse and changing industrial data patterns. Synaptic weights are modified through spike-timing-dependent plasticity rules, enabling the system to learn new relationships and concepts on-the-fly while maintaining structural integrity
Solution Approach 2:
The system changes its internal parameters (synaptic weights, neuronal thresholds, connection strengths) continuously through online learning from sparse industrial data. Rather than relying on fixed pre-defined vocabularies, the neuromorphic system adapts its parameters to match the actual data distributions and relationships observed in dynamic industrial automation systems, achieving both stability through learned structures and versatility through continuous parameter adaptation
4Measurement precision
If external data processing and training infrastructure are used, then comprehensive learning is achieved, but latency increases and energy efficiency decreases for edge deployment
Solution Approach 1:
The patent introduces neuromorphic hardware as an intermediary between data sources and decision-making systems at the edge. This intermediary performs learning and inference locally using spiking neural networks, eliminating the need to transfer data to external processing infrastructure. The intermediary maintains high learning quality through biologically-inspired efficient computations while reducing latency by processing data in real-time at the source, and improves energy efficiency compared to cloud-based offline training approaches
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
Neuromorphic hardware for processing a knowledge graph (KG) represented by observed triple statements (T), with a learning component (LC), consisting of an input layer containing node embedding populations (NEP) of neurons (N), with each node embedding populations (NEP) representing an entity contained in the observed triple statements (T), and an output layer, containing output neurons (SDON) configured for representing a likelihood for each possible triple statement, and modeling a probabilistic, sampling-based model derived from an energy function, wherein the observed triple statements (T) have minimal energy, and with a control component (CC), configured for switching the learning component (LC) into a data-driven learning mode, configured for training the component (LC) with a maximum likelihood learning algorithm minimizing energy in the probabilistic, sampling-based model, using only the observed triple statements (T), which are assigned low energy values, into a sampling mode, in which the learning component (LC) supports generation of triple statements, and into a model-driven learning mode, configured for training the component (LC) with the maximum likelihood learning algorithm using only the generated triple statements, with the learning component (LC) learning to assign high energy values to the generated triple statements.