Device and Method for Similarity Assessment

Simple SummaryContent extracted from patent full text and abstract with AI.

This invention is a hardware device and method for computing "attention" — a key operation in modern AI models like Transformers — using analog electronic circuits instead of conventional digital processors. It works by performing vector-matrix multiplications whose results are output as electrical signals, then converting each signal into a time-delayed pulse where a larger signal produces a later-occurring pulse. A third circuit then derives the attention score from the pattern of these time-dependent pulses. A second vector-matrix multiplication can also be performed to support the full attention mechanism.

Use CasesContent extracted from patent full text and abstract with AI.

• Accelerating Transformer-based neural network inference (e.g., large language models or vision transformers) directly in neuromorphic or analog hardware.
• Deploying energy-efficient AI attention mechanisms in edge devices such as smartphones, IoT sensors, or embedded systems where power budgets are tight.
• Implementing spiking neural network (SNN) architectures that require biologically inspired, pulse-based computation for attention scoring.
• Building dedicated AI inference chips for data centers that need to reduce the energy cost of running attention-heavy models at scale.
• Enabling real-time natural language processing or computer vision in autonomous vehicles or robotics where low latency and low power are critical.

BenefitsContent extracted from patent full text and abstract with AI.

• Replaces power-hungry digital multiply-accumulate operations with analog signal processing, significantly reducing energy consumption per attention computation.
• Encodes computation results as time-delayed pulses, allowing the hardware to exploit temporal dynamics of circuits rather than requiring high-precision digital arithmetic.
• Enables in-memory or near-memory computation of vector-matrix multiplications, reducing data movement bottlenecks compared to conventional von Neumann architectures.
• Supports the full attention mechanism (including a second vector-matrix multiplication) within a compact, purpose-built circuit, reducing chip area and system complexity.
• Naturally maps onto neuromorphic hardware paradigms, making it compatible with emerging spiking neural network ecosystems and bio-inspired computing platforms.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

The invention relates to a method for determining an attention comprising the steps: performing a first vector-matrix multiplication such that each value of the vector resulting from the vector-matrix multiplication is generated in the form of an electrical signal; converting each generated electrical signal into a pulse, the occurrence of which depends on the magnitude of the respectively generated signal, wherein a pulse occurs later the larger the signal is; determining an attention from the pulses. A second vector-matrix multiplication can be performed. A device for performing a method comprises a first electrical circuit which is configured such that a vector-matrix multiplication is performed and the result of the vector-matrix multiplication is output in the form of electrical signals. The device comprises a second electronic circuit which is configured such that each electrical signal is converted into a time-dependent electrical pulse. The device comprises a third electronic circuit which is configured such that it determines an attention from the time-dependent electrical pulses. A second vector-matrix multiplication can be performed with the device.