Improving Peak-picking Using Multiple Time-step Loss Functions.

The majority of state-of-the-art methods for music infor-mation retrieval (MIR) tasks now utilise deep learningmethods reliant on minimisation of loss functions such ascross entropy. For tasks that include framewise binaryclassification (e.g., onset detection, music transcription)classes are derived from output activation functions byidentifying points of local maxima, or peaks. However, theoperating principles behind peak picking are different tothat of the cross entropy loss function, which minimises theabsolute difference between the output and target valuesfor a single frame. To generate activation functions moresuited to peak-picking, we propose two versions of a newloss function that incorporates information from multipletime-steps: 1)multi-individual, which uses multiple indi-vidual time-step cross entropies; and 2)multi-difference,which directly compares the difference between sequentialtime-step outputs. We evaluate the newly proposed lossfunctions alongside standard cross entropy in the popularMIR tasks of onset detection and automatic drum tran-scription. The results highlight the effectiveness of theseloss functions in the improvement of overall system ac-curacies for both MIR tasks. Additionally, directly com-paring the output from sequential time-steps in the multi-difference approach achieves the highest performance.