Intelligent Autonomous Systems

ML approaches are based on two steps: (i) training based on data for the model discovery and (ii) the inference phase, where new data and the model training provide predictions. The training phase often requires large datasets and is computing-intensive. ML algorithms used in decision-making are classified by feedback: supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning.

The Collaborative Learning inspiration source is the student learning context [1]. The student cognitive learning in a group and the problem resolution mode suggest knowledge sharing, consensus, and relationships between different domains [2]. The Federated Learning approach emerged from Google in 2017 as a decentralized decision-making system where selected edge devices deliver collaborative learning to a central server on the cloud using an aggregation algorithm [3], as in Figure 1.

Intelligent Autonomous Systems are IoT devices with computation capability enough to execute machine learning algorithms and take decision-making without constant human supervision but with human collaboration in the loop. The Human-in-the-Loop (HiTL) techniques [4] might be deployed to improve these features.

Figure 1 - Federated Learning Model

Federated Learning training algorithm

The training algorithm follows a data flow similar to the stochastic gradient computing process. First, the devices process their local data, computing hyperparameters based on a loss function, and it sends results without sharing raw data with other partners. Second, the cloud server computes an average from the significant local model in a global hyperparameter with an aggregation algorithm. Finally, the central server updates the local model of devices concerning its global parameter. This recursive processing occurs until achieving high accuracy and the lowest loss function error.

However, this mechanism is heterogeneity sensitive and suffers from non-iid data, i.e., the data is not independent and identically distributed across the devices. Thus, the input data exhibit a bias that needs to be fixed. This project studies how to mitigate these problems and handles the current open issues.

References

[1] T. Okamoto, M. Kayama, and A. Cristea, “Collaborative learning support knowledge management for asynchronous learning networks,” in Proceedings IEEE International Conference on Advanced Learning Technologies, p. 490–491, IEEE Computer Society, Aug. 2001.

[2] C. Robert, Y. Kao, and A. Lepage, “Characterizing Digital Cognitive and Knowledge Strata for Knowledge Management in a Collaborative Learning Environment,” in 2009 Computation World: Future Computing, Service Computation, Cognitive, Adaptive, Content, Patterns, p. 505–510, IEEE Computer Society, Nov. 2009.

[3] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-Efficient Learning of Deep Networks from Decentralized Data,” in Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (A. Singh and J. Zhu, eds.), vol. 54 of Proceedings of Machine Learning Research, (Fort Lauderdale, FL, USA), p. 1273–1282, PMLR, 20–22 Apr 2017.

[4] Fernandes, J., Raposo, D., Armando, N., Sinche, S., Silva, J. Sá, Rodrigues, A., Pereira, V., Oliveira, H. Gonçalo, Macedo, Luís, Boavida, F.: ISABELA – A Socially-Aware Human-in-the-Loop Advisor System , Online Social Networks and Media 16, Elsevier Publishers B. V., 100060, March 2020