Device and Method for Similarity Assessment

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

The patent describes a device and method for efficiently performing similarity evaluations, specifically for use in attention mechanisms like those found in modern machine learning models (e.g., Transformers). The proposed hardware uses capacitors to temporarily store 'key' values and comparators to compare incoming queries against these stored keys, outputting an analog similarity value. This setup allows for the rapid and energy-efficient computation of similarity (a critical operation for attention algorithms), using principles inspired by content addressable memory (CAM) and analog hardware processing.

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

• Speeding up attention computations in deep learning models and neural networks, such as Transformers used in natural language processing and computer vision.
• Hardware acceleration of AI models in edge devices, mobile phones, or IoT devices where energy efficiency is crucial.
• Memory chips or integrated circuits designed for associative search, enabling faster information retrieval based on content similarity rather than address.
• Real-time filtering, pattern recognition, or anomaly detection systems.
• Efficient similarity search engines for high-speed data analytics, recommendation systems, or cybersecurity applications.

BenefitsContent extracted from patent full text and abstract with AI.

• Significantly reduces energy consumption and increases speed of similarity calculations compared to conventional digital approaches.
• Enables more compact, scalable, and cost-effective hardware implementations for machine learning accelerators.
• By performing operations directly in memory (in-memory computing), data movement bottlenecks are minimized, leading to higher overall system efficiency.
• Supports dynamic and frequent updates of stored key values, making it suitable for real-world, rapidly changing workloads.
• Analog similarity outputs offer higher resolution and enable nuanced matching beyond binary/digital comparisons.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

The invention relates to a device for carrying out a similarity evaluation for an attention mechanism with: a capacitor (C1) for storing a key (K); an input for a query (Q); a comparator (Com) for comparing a query (Q); a key (K) stored in the capacitor (C1), wherein the comparator (Com) outputs an analogue similarity value (S) as the result of the comparison. The invention also relates to a method for carrying out a similarity evaluation using a device claimed, wherein keys (K) are stored in capacitors (C1), a query (Q) is then compared with the stored keys (K) by comparators (Com) in order to obtain a similarity. Similarity evaluations can be carried out quickly and efficiently.