Monkey-to-human transfer of brain-computer interface decoders

Abstract

Intracortical brain-computer interfaces (iBCIs) enable paralyzed persons to generate movement, but current methods require large amounts of both neural and movement-related data to be collected from the iBCI user for supervised decoder training. We hypothesized that the low-dimensional latent neural representations of motor behavior, known to be preserved across time, might also be preserved across individuals, and allow us to circumvent this problem. We trained a decoder to predict the electromyographic (EMG) activity for a “source” monkey from the latent signals of motor cortex. We then used Canonical Correlation Analysis to align the latent signals of a “target” monkey to those of the source. These decoders were as accurate across monkeys as they were across sessions for a given monkey. Remarkably, the same process with latent signals from a human participant with tetraplegia was within 90% of the with-monkey decoding across session accuracy. Our findings suggest that consistent representations of motor activity exist across animals and even species. Discovering this common representation is a crucial first step in designing iBCI decoders that perform well without large amounts of data and supervised subject-specific tuning.

Artificial neural network for brain-machine interface consistently produces more naturalistic finger movements than linear methods

“…Research into data augmentation strategies like noise injection and new training paradigms like continual and multitask could improve tcFNN's robustness and improve cross-context performance (47)(48)(49)(50). Two recent studies have shown that incorporating transfer learning into decoder training improves stability over time and allows for cross-species decoding (51,52).…”

Neural Data Transformer 2: Multi-context Pretraining for Neural Spiking Activity

“…Next, given reports of monkey to human transfer [54], we assess whether monkey data in either pretraining or decoder preparation improves decoding (Table 2, rows 4-7). We find that monkey data, however incorporated, reduces offline decoding performance (row 4-7 < 1).…”