ANC Workshop - Xinyu Jiang, Nina Kudryashova *MOVED to 50 George Square*

Tuesday, 19th September 2023 in 50 George Square (G.02)

Plug-and-Play Myoelectric Control - Xinyu Jiang

Abstract: Machine learning for myoelectric control is a central research topic in human-machine interactions. Previous models suffered from high computational complexity, sensitivity to noises, lack of explainability, and performance degradation in long-term applications. Moreover, most machine learning models require a large amount of data from each user to learn the user-specific characteristics of electromyography (EMG) data. In a series of our recent work, we demonstrated how the simple and off-the-shelf random forest-based model evolved step by step into a promising model, that is explainable, robust to noises, robust to inter-day signal variations, computationally efficient, pre-trainable, and can be tuned using a small dataset for a new user. These findings provide a new solution to develop a plug-and-play myoelectric control model.

Uncovering movement planning and corrections: weak behaviour supervision for latent dynamics is all you need - Nina Kudryashova

Abstract:  Throughout evolution, the mammalian brain has developed circuitry that excels at planning movements and adjusting motor execution in response to changes in unpredictable environments (Zador et al., Nat Comm 2023). Before the movement starts, the motor cortical areas prepare for the movement by setting their activity to an optimal initial state for the subsequent movement execution (Kao et al., Neuron 2021). From this initial state, the neural population dynamics unfold to control the motor execution phase and coordinate the sequential activation of muscles. These latent dynamics are known to explain a large fraction of neural variability. However, our recent research has revealed that these dynamics can fail to capture the unplanned movement corrections.

In this talk, I will show that a model of non-autonomous latent neural dynamics that is weakly supervised with behaviour achieves the top performance in behaviour reconstruction. This is demonstrated on a prominent public benchmark within this field, the Neural Latents Benchmark, using a variety of hand-reaching movement datasets.

Finally, I will discuss more biologically plausible variants of utilising behaviour that account for an action-perception loop and can capture both the initially planned movement and the movement correction.

Event type: Workshop

Date: Tuesday, 19th September

Time: 11:00

Location: 50 George Square (G.02) 

Speaker(s): Xinyu Jiang, Nina Kudryashova

Chair/Host: Oksana Sorokina