Open Access
Review
Table 1
Bottleneck identification in the inference process across different tasks
| Bottleneck location | Bottleneck description | Typical tasks | Related research |
|---|---|---|---|
| Modality encoder | Token explosion, high-dimensional feature heterogeneity | VQA, feature extraction, Video preprocessing | [51-54] |
| LLM backbone | Memory-intensive KV cache, massive parameter redundancy | Video captioning, language-guided generation | [10, 11, 55, 56] |
| Cross-modal connector | Alignment overhead, high memory-access ratio | Multimodal retrieval, content synthesis | [57-61] |
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.