FedPEAT: Convergence of Federated Learning, Parameter-Efficient Fine Tuning, and Emulator Assisted Tuning for Artificial Intelligence Foundation Models with Mobile Edge Computing

Abstract

The emergence of foundation models, including language and vision models, has reshaped AI's landscape, offering capabilities across various applications. Deploying and fine-tuning these large models, like GPT-3 and BERT, presents challenges, especially in the current foundation model era. We introduce Emulator-Assisted Tuning (EAT) combined with Parameter-Efficient Fine-Tuning (PEFT) to form Parameter-Efficient Emulator-Assisted Tuning (PEAT). Further, we expand this into federated learning as Federated PEAT (FedPEAT). FedPEAT uses adapters, emulators, and PEFT for federated model tuning, enhancing model privacy and memory efficiency. Adapters adjust pre-trained models, while emulators give a compact representation of original models, addressing both privacy and efficiency. Adaptable to various neural networks, our approach also uses deep reinforcement learning for hyper-parameter optimization. We tested FedPEAT in a unique scenario with a server participating in collaborative federated tuning, showcasing its potential in tackling foundation model challenges.

Figures (7)

• Figure 1: Intersection of Federated learning (FL), Parameter-Efficient Fine-Tuning (PEFT), and Emulator-Assisted Tuning (EAT). Here we illustrate the intersection of FL, PEFT, and (EAT). The main contribution of our current paper is to introduce Federated Parameter-Efficient Emulator-Assisted Tuning (FedPEAT), as a convergence of EAT, PEFT, and FL, while EAT and Parameter-Efficient Emulator-Assisted Tuning (PEAT) are also terms coined by our paper.
• Figure 2: FedPEAT with Adaptive control overview. This figure shows how the Adaptive control orchestrator makes decisions on important parameters, such as device selection, emulator compression parameter, transmission bandwidth and power to facilitate the FedPEAT process.
• Figure 3: Emulator-Assisted Tuning generalized to three cases. Figure illustrates how the neural network structures at the server and local devices differ in each case. Case 1 represents our proposed FedPEAT framework. Case 2 represents the integration of Federated Learning and PEFT. Case 3 represents a traditional Federated Learning scenario.
• Figure 4: Our proposed SABPPO algorithm and architecture. Figure illustrates the underlying actor and critic architecture, their interaction with the environment and model update process.
• Figure 5: Comparison between FL and FedPEAT, and Comparison between Adaptive control algorithms. Figure \ref{['fig:fl_delay']}, \ref{['fig:fl_emuxchange']}, \ref{['fig:fl_perplexity']} illustrates the performance difference between FL and FedPEAT with regards to delay, emulator exchange count, and perplexity, respectively. Figure \ref{['fig:traintime']} illustrates the time taken for model training for each adaptive control algorithm. Figure \ref{['fig:fed_delay']}, \ref{['fig:fed_emuxchange']}, \ref{['fig:fed_perplexity']}, \ref{['fig:fed_reward']} illustrate the performance of each adaptive control algorithm in across the training process, in terms of log(delay), emulator exchange count, perplexity, and reward.