TEACHING Platform for Human-Centric Autonomous Applications: Design and Overview

Valerio De Caro, Christos Chronis, Massimo Coppola, Vincenzo Lomonaco, Claudio Gallicchio, Konstantinos Tserpes, Davide Bacciu

Sure, here's a blog-style summary of the academic paper:

TL;DR

The TEACHING platform is a novel microservice-based framework designed to foster "Humanistic Intelligence" by seamlessly integrating humans with Cyber-Physical Systems of Systems (CPSOS). It enables dynamic, AI-powered applications that adapt to human users by treating them as first-class citizens, closing the feedback loop between human and machine.

What Problem Does This Paper Address?

The paper tackles the challenge of creating truly human-centric autonomous systems. Traditional human-computer interaction often treats humans as external inputs or outputs, rather than integral components. This leads to systems that may not fully understand or adapt to human needs, preferences, and states. The authors aim to bridge this gap by developing a platform that facilitates a symbiotic relationship between humans and AI-enabled CPSOS, where each mutually empowers the other. This is particularly relevant in complex, pervasive environments where systems need to be flexible, adaptable, and trustworthy.

Methods

The core of the TEACHING platform is its microservice architecture. This allows for modularity, scalability, and dynamic composition of functionalities. Key components include:

  • Message Broker (RabbitMQ): Facilitates communication between microservices, enabling dynamic data workflow definitions.
  • Docker and Docker-Compose: Used for containerization, ensuring efficient lifecycle management, isolation, and portability of services.
  • CI/CD Pipeline (GitHub Actions): Automates the building, testing, and deployment of platform modules, streamlining development.
  • AI-Toolkit: A collection of AI modules (sequence learning, federated learning, stress prediction, etc.) that can be integrated into applications.
  • Model Transfer Service (MTS): Ensures secure, end-to-end encryption of model weights during transmission, crucial for federated learning and data privacy.
  • Data Stream Processing (DSP) with WindFlow: Enables high-speed processing of streaming data from sensors, with efficient execution on embedded devices.
  • External Services Integration: Allows for seamless aggregation of data and knowledge across distributed platforms, enhancing privacy and performance.

The platform's design emphasizes programmability, allowing developers to define data workflows using an abstract language, with minimal code intervention.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Humanistic Intelligence (HI): The central concept is treating humans and AI as intertwined entities, where AI augments human capabilities. This moves beyond simple human-in-the-loop systems to a more symbiotic relationship.
  • Humans as First-Class Citizens: The platform is designed to actively integrate human reactions and states into the system's feedback loop, rather than treating them as mere external data points.
  • Dynamic Workflow Composition: The microservice architecture, coupled with a message broker, allows for the creation of highly flexible and dynamic data workflows that can adapt to changing requirements and environments.
  • Closing the Feedback Loop: By enabling mutual feedback between humans and AI-enabled CPSOS, the system can dynamically adapt its behavior to better suit the user's state and preferences.
  • Edge-to-Cloud Continuum: The platform supports integration across distributed environments, from edge devices to the cloud, facilitating seamless data and model exchange.

Key Results

The paper primarily focuses on the design and architecture of the TEACHING platform and demonstrates its application through a validation use case in autonomous driving.

  • Autonomous Driving Use Case: The platform was used to develop an application where a car's driving profile is adjusted to minimize driver stress, based on physiological data. This involves:
  • Monitoring and quantifying human stress levels.
  • Using a personalization model (trained federatedly) to adapt driving profiles.
  • Integrating with the car's AI system to modify driving behavior.
  • Hardware Platform: The validation utilized an NXP iMX8 Quad Max board, highlighting the platform's ability to handle both real-time processing and AI computations.
  • Development Process: The paper outlines how application providers can define complex applications by selecting and configuring microservices, demonstrating the platform's ease of use for developers.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges the complexity of deploying in pervasive environments with heterogeneous hardware and the challenges of human-CPSOS interaction.
  • While the autonomous driving use case is detailed, specific quantitative performance metrics (e.g., stress reduction percentages, latency improvements) are not extensively presented.
  • Future Work:
    • Transpiling Mechanism: The authors aim to develop a mechanism that allows entire applications to be defined using simple Python code.
  • Platform GUI: A graphical user interface (GUI) is planned to enable application building through the visual definition of computational graphs.
  • Further exploration of the challenges in human-CPSOS interaction and pervasive environment deployment.

Summary updated

BibTeX
@inproceedings{caro2024teaching,
  author = {Valerio De Caro and Christos Chronis and Massimo Coppola and Vincenzo Lomonaco and Claudio Gallicchio and Konstantinos Tserpes and Davide Bacciu},
  title = {{TEACHING Platform for Human-Centric Autonomous Applications: Design and Overview}},
  booktitle = {HPDC},
  year = {2024},
  doi = {10.1145/3625549.3658813},
}
Figure