A.1 TinyML for IoT networks- Design of pull-based communication framework

[Motivation and Context] In 6G communication systems, especially for 6G Internet of Things (IoT) networks, Artificial Intelligence (AI)/Machine learning (ML) models play a key role in realizing intelligence at the edge. Energy-efficient connectivity is one of the essential key requirements for 6G IoT Networks. The attractive approach to realize this is to design the communication protocol in such a way that the IoT device transmits only the relevant data at the right timing. In order to realize this vision, we have introduced the pull-based communication system for 6G IoT networks integrating TinyML model, as illustrated in the figure below.

Fig. 1 Illustration of pull-based communication framework integrating TinyML models [ref-A1]

[Ref1] M. Thorsager, V. Croisfelt, J. Shiraishi and P. Popovski, "EcoPull: Sustainable IoT Image Retrieval Empowered by TinyML Models," GLOBECOM 2024 - 2024 IEEE Global Communications Conference, Cape Town, South Africa, 2024, pp. 5066-5071, doi: 10.1109/GLOBECOM52923.2024.10901782. 

[Basic Idea] As shown in the figure 1, our proposed communication framework for 6G IoT networks, in which the Edge server (ES) equipped with a foundation model is interested in retrieving the specific types of data by sending a semantic query. Our idea is to introduce two types of small ML models into IoT devices so that ES can retrieve only the relevant latent from the IoT devices: 1) Tiny behavior model and 2) Tiny image compressor model. The behavior model decides whether to transmit data based on the received query by evaluating the data's relevance to the current task at the ES. On the other hand, the image compressor model encodes the image into the latent space to reduce communication costs. As illustrated in Fig. 1, we consider a pull-based communication framework. When the ES wants to collect the specific data, it first sends a semantic query. The IoT device receiving a semantic query starts processing all stored data--it first checks the relevance of the stored image by comparing the feature vector of each image with the received query feature vector. If the matching score of an image is higher than a predetermined threshold, it then compresses its data by using an image compressor model and stores it in the query. After processing, each IoT device transmits relevant data over a shared medium by random access. The ES then reconstructs the original image based on the received latency by using generative AI.  By applying this framework, we can allocate the limited wireless resources to only the important data, reducing the overall energy consumption of IoT devices. 

[Results] By caredully tuning the semantic filtering size and latent transmission size, our proposed framwoerk can save wasteful energy consumption compared with the simple offloading scehme while providing high utility.

A2. Energy efficient Continual Learning framework for IoT Networks

B.1) Content-based Wake-up for Timely IoT Sensing Retrieval

[Background] This work focuses on how to collect informative data with high energy efficiency and high information freshness in order to serve the Industrial Internet of Things applications, such as remote control of mobile actuators. Specifically, this work considers how the sink can collect k highest readings among the deployed sensors with high energy efficiency while providing high information freshness at the deadline time via wake-up radio. One of the promising approaches is content-based wake-up radio[Ref-B1-1 ][Ref-B1-2], in which the information of the specific content, such as the threshold for the observed data, is encoded into the wake-up signal. The wake-up receiver detecting this signal checks whether the current observations satisfy the condition specified by the wake-up signal. If so, it activates the main radio and transmits the current observations to the sink. 

Fig. B1-1. An example of the top-k query employing CoWu, where k = 2. (a): The sink transmits CoWu signal at the early time against the deadline. (b): The sink transmits CoWu signal at the late time against the deadline. (c) An example on how the timing and the threshold of CoWu signal affects the accuracy of a top-k set at the deadline. 

Context:

In communication, two types of communication modes are considered: push-based communication and pull-based communication. The push-based communication follows an event-driven approach, excelling in scenarios that require immediate reporting and response, such as anomaly detection. Pull-based communication is more effective in scenarios where the edge or cloud runs a comprehensive AI model that can impose global orchestration of the data collection and selectively retrieve data, thereby reducing irrelevant data traffic [REF-C1]. 

E. Neuromorphic Communication

Contention-free pull and contention-push (CFC-pull/push) [REF-C2]

• Pull unicast scheduling:

  • enabling the retrieval of sensor readings from specific devices at specific times without collisions

• Push contention:

  • Devices transmit locally relevant data by contending for the channel with other push-based traffic.

Fig. D1 Illustration of proposed wireless memory activation and approximation [Ref-D1]

 [Ref-D1] Junya Shiraishi, Shashi Raj Pandey, Israel Leyva-Mayorga, Petar Popovski, “Wireless memory approximation for energy-efficient task-specific IoT data retrieval,” in Proc. IEEE International Conference on Communications 2026, 2026.