58
0
0
2021-06-05

Special Issue on Recent Progress in Autonomous Machine Learning Submission Date: 2021-07-01 Autonomous Machine Learning (AML) refers to a learning system having flexible characteristic to evolve both its network structure and parameters on the fly. It is capable of initiating its learning process from scratch with/without a predefined network structure while its knowledge base is automatically constructed in real-time. AML is built upon two fundamental principles: one-pass learning strategy and self-evolving network structure. The former one reflects a situation where a data point is directly discarded once learned to assure bounded memory and computational burdens while the latter lies in the self-reconfiguration aptitude of AML where its network size can increase or reduce in respect to varying data distributions. AMLs have been proven to be useful in handling real-time data streams where a learning system confronts never-ending information flow which does not follow static or predictable data distributions rather drifting overtime with different types, magnitudes and types. Variants of AMLs are capable of quickly reacting to those drifting distributions regardless of how slow, fast, sudden, gradual, cyclic changing distributions might be while retaining computationally light characteristics. In addition, the AMLs have grown into various application domains not only limited to regression, classification, clustering but also control and reinforcement learning. In a nutshell, it is enabled by the fact that AMLs aim to balance between stability and plasticity of a learning system.


Recent challenges in machine learning renders innovation of AMLs urgently needed. The advent of deep learning technologies is a concrete example. Existing DNNs mostly rely on a static and offline learning principle limiting its feasibility in the streaming environments. On the other hand, DNNs are well-known for its feature learning power being able to handle unstructured problems with large input dimension and target classes. The network structure of DNNs are difficult to evolve because of the absence of local and spatial contexts. The multi-layer nature of DNNs complicate the self-evolving strategy. Insertion of a new layer definitely leads to the catastrophic forgetting problem. Another research opportunity of AMLs is identified in the context of lifelong/continual learning where the goal is not only to adapt to changing environments but also to actualize a lifelong learning agent with knowledge retention property. That is, a learning agent must not suffer from the catastrophic forgetting problem when adapting to a new context. The fact that AMLs are normally designed in the local learning environment should be useful for this purpose. Only relevant knowledge is stimulated by new tasks while others remain silent. The application of AMLs in the transfer learning domain deserves in-depth study. Unlike traditional AML involving only a single stream, the case of multi-streams remains an open issue. The main goal of this problem is to create a domain-invariant network handling both source stream and target stream equally well. The challenge of this topic is evident in the covariate shift problem between source stream and target stream. As with the single stream case, the concept drift occurs here in each stream in different time periods.


This special issue aims to bring together recent research works of AMLs. Particular interest lies in the integration of AMLs in handling advanced issues of machine learning as abovementioned. We solicit original works that have not been published nor under consideration in other publication venues.


Topic of Interest


The topic of interest includes the following but not limited to


Novel network architecture of AMLs

AMLs to handle unstructured problems such as texts, videos, speech, etc.

AMLs to handle weakly supervised learning problem.

AMLs to handle semi-supervised learning problem.

Active learning for AMLs.

AMLs to handle continual learning problem.

AMLs to handle multi-stream problem.


Important Dates


Manuscript Submission Deadline: July 1st, 2021

First Round of Reviews: September 30th, 2021

Revised Paper Submission: December 10th, 2022

Second Round of Reviews: February 1st, 2022

Expected Publication Date: May 1st, 2022


Submission Instruction


To be considered in this special issue, author should select VSI: Recent Progress in AML in the Elsevier editorial system.


Guest Editors


Mahardhika Pratama

email: mpratama@ntu.edu.sg

School of Computer Science and Engineering

Nanyang Technological University

Singapore


Edwin Lughofer

email: edwin.lughofer@jku.at

Department of Knowledge-Based Mathematical Systems

Johannes Keppler University

Austria


Plamen P. Angelov

email: p.angelov@lancaster.ac.uk

School of Computing and Communications

Lancaster University

UK

登录用户可以查看和发表评论, 请前往  登录 或  注册


SCHOLAT.com 学者网
免责声明 | 关于我们 | 用户反馈
联系我们: