Semantic Meta-Split Learning: A TinyML Scheme for Few-Shot Wireless Image Classification

Eslam Eldeeb,Mohammad Shehab,Hirley Alves,Mohamed-Slim Alouini

2024-09-03

Abstract:Semantic and goal-oriented (SGO) communication is an emerging technology that only transmits significant information for a given task. Semantic communication encounters many challenges, such as computational complexity at end users, availability of data, and privacy-preserving. This work presents a TinyML-based semantic communication framework for few-shot wireless image classification that integrates split-learning and meta-learning. We exploit split-learning to limit the computations performed by the end-users while ensuring privacy-preserving. In addition, meta-learning overcomes data availability concerns and speeds up training by utilizing similarly trained tasks. The proposed algorithm is tested using a data set of images of hand-written letters. In addition, we present an uncertainty analysis of the predictions using conformal prediction (CP) techniques. Simulation results show that the proposed Semantic-MSL outperforms conventional schemes by achieving 20 % gain on classification accuracy using fewer data points, yet less training energy consumption.

Computer Vision and Pattern Recognition,Machine Learning,Image and Video Processing

What problem does this paper attempt to address?

### Problems Addressed by the Paper

This paper aims to address several key challenges in wireless image classification, particularly in the context of few-shot scenarios, focusing on efficient, low-energy, and privacy-preserving wireless image classification. Specifically:

1. **Semantic Communication Framework**: A TinyML framework combining split-learning and meta-learning, called "Semantic Meta-Split Learning" (Semantic-MSL), is proposed for wireless image transmission and classification.

2. **Computational Complexity and Privacy Protection**: Split learning is used to reduce the computational load on terminal devices and ensure data privacy, as only processed data (smashed data) is transmitted to the aggregator.

3. **Data Availability and Rapid Training**: Meta-learning is utilized to overcome data scarcity issues and accelerate the learning process for new tasks by leveraging the experience from similar tasks.

4. **Experimental Validation**: The framework is tested using a handwritten alphabet image dataset and evaluated for prediction uncertainty using conformal prediction (CP) techniques. Experimental results show that compared to traditional split learning (Semantic-SL) and deep neural networks (DNN), Semantic-MSL significantly improves classification accuracy, conformal prediction uncertainty, and energy consumption.

In summary, the main goal of this paper is to achieve efficient, low-energy, and privacy-preserving wireless image classification in few-shot scenarios. By combining meta-learning and split learning, the proposed method addresses the high computational complexity, data insufficiency, and privacy leakage issues present in existing methods.