OPERATION OF BIDIRECTIONAL RECURRING NEURAL NETWORKS IN HARDWARE
Patent Information
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- IMAGINATION TECH LTD
- Filing Date
- 2022-06-29
- Publication Date
- 2026-06-10
AI Technical Summary
Implementing Bidirectional Recurrent Neural Networks (BRNNs) in hardware is challenging due to their dynamic nature, as each cell's operations depend on both forward and backward states, creating a circular dependency that is difficult to execute efficiently on accelerators designed for static neural networks.
The approach involves unrolling each BRNN cell into a pair of forward and backward recurrent neural networks over a predetermined number of timesteps, transforming the dynamic graph into a static graph suitable for implementation on conventional accelerators.
This method allows BRNNs to be efficiently executed on hardware by converting them into a static graph, enabling the same accelerators to handle both recursive and non-recursive neural networks, thereby extending their utility.