A serializing device for convolutional filtering
The serializer optimizes CNN hardware implementations by converting input feature maps into serialized vectors for parallel processing, addressing computational and memory management challenges, enhancing efficiency and integration in resource-constrained environments.
WO2026139438A1PCT designated stage Publication Date: 2026-07-02INSTITUT MINES TELECOM TELECOM BRETAGNE
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- INSTITUT MINES TELECOM TELECOM BRETAGNE
- Filing Date
- 2025-12-19
- Publication Date
- 2026-07-02
Smart Images
Figure EP2025088618_02072026_PF_FP_ABST
Abstract
There is provided a serializer (3) configured to receive an input feature map represented by a Q-dimensional tensor and to convert the received input feature map into a series of 1D vectors, each corresponding to a convolutional patch. The serializer comprises a memory device (300) comprising a plurality of ports and of memory banks adapted to store data words, the ports comprising a least a write port and at least two read ports. The serializer (3) is adapted to perform write operations consisting in writing a number of 1D vectors to a number of banks of said memory device (300) using the writing port, while being adapted for reading in parallel convolutional patches corresponding to stored 1D vectors from banks of the memory device (300) using the at least two reading ports, the extracted convolutional patches being fed to a convolutional filtering device (90) implementing convolutional filters.
Need to check novelty before this filing date? Find Prior Art