ReScience C Special Journal Issue
Our reproducibility studies, produced as part of a learning course, have been accepted as a special issue in collaboration with ReScience C and are now in editing ahead of publication.
Our reproducibility studies, produced as part of a learning course, have been accepted as a special issue in collaboration with ReScience C and are now in editing ahead of publication.
Shows that a structural load forecasting model cannot separate out the drop in demand caused by a public conservation appeal.
Uses cartographic generalization metrics to measure how downsampling dermoscopy benchmarks strips away lesion structure unevenly across diagnostic classes.
Walks through seven flaws in our own evaluation protocol, from ancestral sequence leakage to train-test contamination, that each gave us a wrong answer before being caught.
Uses token-level predictive entropy to find the hardest line breaks in a poem and has the model stop short of writing them.
Measures the meter and word choice a language model inherits from its pretraining corpus and argues the resulting voice is recall of that corpus rather than writing of its own.
Derives spectral diagnostics that predict when mode truncation makes a Fourier neural operator surrogate lose accuracy on reaction-diffusion systems.
Uses causal responsibility attribution to work out which individual macro placements are responsible for later wirelength and timing problems.