Search
DeepSeek Teams Up with Peking University to Introduce Engram Architecture, Overcoming Transformer's Memory Limitations
3 week ago /
Read about 0 minute
Author：小编
According to a report by 36Kr, DeepSeek has joined forces with Peking University to release a new research paper, unveiling the "Engram" module. Spearheaded by researchers such as Liang Wenfeng, this innovative module is designed to tackle the inherent limitation of the Transformer model—its inability to natively retrieve knowledge. The Engram module incorporates a "conditional memory" sparse axis, leveraging cutting-edge hashed N-gram embedding techniques. This approach enables deterministic knowledge retrieval with an approximate time complexity of O(1), substantially boosting the model's proficiency in tasks involving knowledge recall, reasoning, coding, and mathematical computations.
-
C114 Communication Network -
Communication Home
7 X 24 Track global technological trends
