Muscle-signal pretraining benefits shrink with unseen test participants
William Raftery and Dario Farina have tested three pretrained muscle-signal models using gesture datasets. Gains observed when training and testing share a recording session largely fade when the test uses entirely different people. The preprint compares pretrained networks with the same architectures trained from scratch, and finds simpler methods competitive in datasets with relatively few participants.
Artificial Intelligence··Midday
New people change the test
William Raftery and Dario Farina released their muscle-signal evaluation on 5 November. They study surface electromyography, which records muscle activity through the skin, using three models: TinyMyo, PhysioWave and NeuroRVQ. The tasks classify hand gestures. A model may train and test on material from one person’s recording session, or face test participants who never appeared in its training data. The researchers compare both arrangements to examine how well pretraining transfers between people.[1]
Separating pretraining from architecture
The study uses three gesture datasets, with 10, 612 and 36 participants respectively. Each pretrained network is compared with the same architecture initialized randomly and trained from scratch. Another arrangement freezes the pretrained backbone and trains only its output layer. The experiments also examine preprocessing choices and repeat training with different random starts. Comparisons include classical approaches such as support vector machines and random forests, alongside a simple convolutional network.[1]
Gains shrink across participants
Pretraining generally helps when training and testing share a session. Its benefit is small or absent for unseen participants in the two smaller datasets, where classical methods can match or exceed the foundation models. In the larger dataset, the benefit also decreases but persists. The authors recommend testing on separate participants and including a comparison with training from scratch. Their findings concern gesture classification in these datasets; the experiments do not establish clinical benefits for patients or people using prostheses.[1]
Related columns
For more information on this topic, you can read the related columns.
