[HUGGINGFACE]score: 0.42
Recurrent CNN Trades Breadth, Depth, and Time via Differentiable Cost Terms
May 23, 2026
A recurrent convolutional network defined on an infinite lattice learns to allocate computation across breadth, depth, and recurrent time steps jointly with task loss via backpropagation. Networks spontaneously take more recurrent steps on occluded inputs, and time usage correlates with human reaction time data.
paper
HOW THIS AFFECTS YOU
●
researcherDifferentiable resource cost terms that produce emergent computational graphs are a novel training technique with potential implications for efficient architecture search and neuroscience-aligned modeling.