Search papers, labs, and topics across Lattice.
The Hong Kong Polytechnic University
5
0
8
A single EEG decoding architecture, DSAINet, achieves state-of-the-art generalizability across diverse tasks and datasets without task-specific tuning, despite having only 77K parameters.
Decoupling memory conditioning from video generation allows for more data-efficient training and better spatial consistency in long-horizon video generation, even when exploring novel scenes.
Fine-tuning the exploration-exploitation balance can dramatically boost LLM reasoning capabilities, as shown by our novel perplexity-guided strategy.
LLMs can navigate complex tool-use scenarios more effectively by organizing tools into specialized agents and aligning high-level planning with agent execution.
By fusing temporal and spatial EEG features layer-by-layer, LI-DSN shatters the "information silo" problem of traditional dual-stream networks, unlocking significant gains in decoding accuracy across diverse BCI tasks.
