Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Convergence Rate Maximization for Split Learning-based Control of EMG Prosthetic Devices

Matea Marinova, Daniel Denkovski, Hristijan Gjoreski, Zoran Hadzi-Velkov, Valentin Rakovic

TL;DR

The paper addresses convergence-rate maximization in Split Learning for EMG-based prosthetic control on resource-constrained wearables. It introduces OCLA, a two-phase offline-online cut-layer selection algorithm that uses profile-function pruning and a $\Delta$-based trade-off to identify optimal split points under varying compute and communication conditions, with an online optimality condition $\Delta(n,n+1) < \frac{\beta R}{f_k} < \Delta(n-1,n)$ where $a = \frac{f_s}{f_k}$ and $\beta = \frac{a-1}{a}$. Through a simulated EMG pattern recognition task using a CNN across 10 subjects, OCLA demonstrates faster convergence and higher accuracy than a naive fixed cut-layer approach, with larger gains under greater system variability. This work enables practical SL deployment in EMG prosthetics by coupling model partitioning to dynamic resource profiling.

Abstract

Split Learning (SL) is a promising Distributed Learning approach in electromyography (EMG) based prosthetic control, due to its applicability within resource-constrained environments. Other learning approaches, such as Deep Learning and Federated Learning (FL), provide suboptimal solutions, since prosthetic devices are extremely limited in terms of processing power and battery life. The viability of implementing SL in such scenarios is caused by its inherent model partitioning, with clients executing the smaller model segment. However, selecting an inadequate cut layer hinders the training process in SL systems. This paper presents an algorithm for optimal cut layer selection in terms of maximizing the convergence rate of the model. The performance evaluation demonstrates that the proposed algorithm substantially accelerates the convergence in an EMG pattern recognition task for improving prosthetic device control.

Convergence Rate Maximization for Split Learning-based Control of EMG Prosthetic Devices

TL;DR

The paper addresses convergence-rate maximization in Split Learning for EMG-based prosthetic control on resource-constrained wearables. It introduces OCLA, a two-phase offline-online cut-layer selection algorithm that uses profile-function pruning and a -based trade-off to identify optimal split points under varying compute and communication conditions, with an online optimality condition where and . Through a simulated EMG pattern recognition task using a CNN across 10 subjects, OCLA demonstrates faster convergence and higher accuracy than a naive fixed cut-layer approach, with larger gains under greater system variability. This work enables practical SL deployment in EMG prosthetics by coupling model partitioning to dynamic resource profiling.

Abstract

Split Learning (SL) is a promising Distributed Learning approach in electromyography (EMG) based prosthetic control, due to its applicability within resource-constrained environments. Other learning approaches, such as Deep Learning and Federated Learning (FL), provide suboptimal solutions, since prosthetic devices are extremely limited in terms of processing power and battery life. The viability of implementing SL in such scenarios is caused by its inherent model partitioning, with clients executing the smaller model segment. However, selecting an inadequate cut layer hinders the training process in SL systems. This paper presents an algorithm for optimal cut layer selection in terms of maximizing the convergence rate of the model. The performance evaluation demonstrates that the proposed algorithm substantially accelerates the convergence in an EMG pattern recognition task for improving prosthetic device control.
Paper Structure (6 sections, 2 theorems, 21 equations, 7 figures, 2 tables, 1 algorithm)

This paper contains 6 sections, 2 theorems, 21 equations, 7 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. System Model
  4. Optimal Cut Layer Selection Algorithm
  5. Results
  6. Conclusion

Key Result

Lemma 1.1

Assume that the communication-computation trade-off between layer $n$ and layer $n+1$ is known and specified as in (gain_plus), and that (Tplus) always holds. If $n$ is the optimal cut layer, then the respective communication-computation trade-off is bounded by:

Figures (7)

  • Figure 1: Generic split learning architecture composed of N clients, which contain the first three layers of the neural network; the remaining four layers are assigned to the server
  • Figure 2: Cumulative client-side computational load per layer
  • Figure 3: Per-layer activation size
  • Figure 4: Cumulative client-side model parameters per layer
  • Figure 5: Performance gain of OCLA regarding the coefficients of variations of the transmission rate $R_{cv}$ and the computing speeds' ratio $(1-\beta)_{cv}$; the baseline algorithm consistently selects layer 3
  • ...and 2 more figures

Theorems & Definitions (2)

  • Lemma 1.1
  • Lemma 1.2