P010Filed

Recurrent Neural Network Inference in Ternary Domain

Gradients compressed to three values — 8× less bandwidth across the training cluster.

AU Application
2023900010
Filing Date
1 March 2023
Index Number
P010
Figures
11 figures
Batch / Category
Core 1

Explore the Vision

Discover this technology through five complementary perspectives — from technical architecture to partnership outcomes. Each layer reveals a different aspect of how this innovation creates value.

Gradients compressed to three values — 8× less bandwidth across the training cluster.

What It IS

Technical Vision

The architectural essence — what makes this technology work

Streams of gradient data flowing between training nodes, each stream transforming from a dense river of floating-point numbers into a crystalline lattice of three symbols. The information content survives; the bandwidth cost collapses. A training cluster breathing through a straw and thriving.

1/5
Explore the buyer's journey across 5 perspectives

Abstract

Techniques for ternary quantization of LSTM and GRU architectures, enabling efficient on-device sequence processing with minimal state overhead.

Visual Essence

Streams of gradient data flowing between training nodes, each stream transforming from a dense river of floating-point numbers into a crystalline lattice of three symbols. The information content survives; the bandwidth cost collapses. A training cluster breathing through a straw and thriving.

Visual Family:compression-crystal

Technology Domains

Related Patents

From the compression-crystal visual family