Structural learning under N:M constraints
SCHTs combines dynamic sparse training with fine- and coarse-grained N:M constraints, exploring network structures aligned with hardware computation.
From models to architecture
The official abstract reports accuracy above the dense baseline at 90% sparsity in the evaluated spiking networks. It also explores indexing and computation in the Trapezoid architecture through cycle-accurate simulation.
Why it matters
The work connects the structure a network learns with how it is executed in hardware. It informs model–hardware co-design; system-level results depend on the platform, model and execution environment.
Based on the official conference abstract; consult the paper for experimental configurations and figures.