🇹🇼 Chinese, 25 October 2026, sparse connections
Researchers Explore Efficient AI with Sparse Connections
A team at Tsinghua University is developing a technique to reduce unnecessary computations in AI models.
The researchers, led by Carlo Vittorio Cannistraci, are inspired by the human brain’s efficient communication and task-specific connections.
They are working on a technology called ‘connectivity sparsity’ to let AI models adjust their internal information pathways during learning, preserving useful connections and reducing unnecessary computations.
Initial experiments have shown promising results, with a neural network retaining only 5% of its connections while outperforming a fully connected network in machine translation tasks.