Over-the-Air Split Learning with MIMO-Based Neural Network and Constellation-Based Activation

TL;DR

This paper proposes a communication-efficient over-the-air split learning system using MIMO and constellation-based activation, which reduces costs and improves efficiency by eliminating explicit channel estimation.

Contribution

It introduces a novel MIMO-based split learning framework with trainable precoding and combining, leveraging constellation diagrams as activations to enhance over-the-air computation.

Findings

Significant reduction in system costs due to implicit channel estimation

Improved efficiency through trainable MIMO precoding and combining

Effective use of constellation diagrams as activation functions

Abstract

This paper investigates a communication-efficient split learning (SL) over multiple-input multiple-output (MIMO) communication system. In particular, we mathematically decompose the inter-layer connection of a neural network (NN) to a series of linear precoding and combining transformations using over-the-air computation (OAC), which synergistically form a linear layer in NNs. The precoding and combining matrices are trainable parameters in such a system, whereas the MIMO channel is implicit. The proposed system eliminates the implicit channel estimation through exploiting the channel reciprocity and properly casting the backpropagation process, significantly saving the system costs and further improving the overall efficiency. The practical constellation diagrams are used as the activation function to avoid sending arbitrary analog signals as in the traditional OAC system. Numerical results are illustrated to demonstrate the effectiveness of the proposed scheme.