Fig. 5: Performance analysis of Reminisce’s key designs and query latency impact on ORIN (INT8).
From: Ubiquitous memory augmentation via mobile multimodal embedding system
a Throughput-to-accuracy trade-off with and without Reminisce’s key designs (1, 2, 3). PE refers to pre-exited coarse-grained embeddings without fine-grained upgrading during the query phase. b Performance under different query latency tolerance.
