Master of Science Rohit Agarwal will Wednesday November 12th, 2025, at 12:15 hold his Thesis Defense for the PhD degree in Science. The title of the thesis is:
« Scalable Online Deep Learning for Streaming Data with Variable Feature Spaces: Architectures, Imparting Scalability and Data Transformation »
The surge of streaming data in real-world applications demands real-time modeling, making deep learning essential for handling the vast amounts of information generated. This has given rise to online learning, a branch of AI where data is processed immediately upon arrival without being stored. Traditionally, online learning assumes a fixed input dimension, yet real-world applications often generate a variable number of features, posing a modeling challenge. This thesis defines such variable feature spaces as haphazard inputs. Existing deep learning models in online settings inherently assume fixed input dimensions, which renders them unsuitable for haphazard inputs. This thesis argues that Scalable Online Deep Learning (SODL) models are necessary for effectively handling haphazard inputs, where scalability is defined as the capability of a model to adapt in real time to varying input feature spaces. The primary contribution of this thesis is pioneering the field of SODL for haphazard inputs. This thesis explores three research verticals, leading to various SODL methods. Since no existing deep learning architecture can handle haphazard inputs, the first vertical involves developing new SODL architectures that can effectively model haphazard inputs. Although several state-of-the-art online deep learning models exist, they remain restricted to fixed feature spaces. Thus, the second vertical investigates the possibility of imparting scalability to existing online deep learning models, making them adept at processing haphazard inputs. While both of these verticals emphasize new architectures or architectural modifications, the third vertical focuses on data adaptation, where varying feature spaces are transformed into fixed representations to leverage established deep learning methods without any structural modifications. The first research vertical is addressed through the development of three SODL architectures, namely, Aux-Net, HapTransformer, and packetLSTM. Aux-Net proposes dynamically adjustable layers within a multilayer perceptron, thus enabling the model to scale its architecture at runtime and thereby handle haphazard inputs. HapTransformer introduces a transformer-based model that enforces permutation variance across input features, followed by processing with a decoder capable of accepting variable inputs. The packetLSTM architecture proposes a recurrent neural network framework that allocates one sub-model per feature and facilitates inter-feature knowledge sharing, achieving scalable learning from haphazard inputs. The second research vertical is addressed through the development of Aux-Drop, an SODL concept that imparts scalability to existing multilayer perceptron-based online deep learning models. Aux-Drop functions as a dropout-based regularization mechanism that enables these models to train multiple sub-networks. At each time step, an appropriate sub-network can be selected, thereby allowing the model to effectively manage the variability inherent in haphazard inputs. The third research vertical is addressed through the development of two SODL methods, namely, Haphazard Inputs as Images (HI2) and Variable to Fixed (V2F). HI2 transforms variable feature spaces into fixed-size graphical representations such as bar chart images, facilitating the use of pre-trained vision models within an online learning framework. V2F proposes a data transformation technique that generates fixed-length numerical embeddings from variable feature spaces in real time, enabling traditional deep learning models to process haphazard inputs without any structural modifications. This thesis, therefore, establishes the field of SODL and proposes six methods that effectively address haphazard inputs, demonstrating their applicability to real-world scenarios and paving the way for future research directions.
1st Opponent: Professor Linga Reddy Cenkeramaddi, University of Agder, Norway
2nd Opponent: Associate Professor Dawood Al Chanti, National School of Physics, Grenoble, France
Internal member and leader of the committee: Associate Professor Elisavet Kozyri, Department of Computer Science, UiT
The defence and trial lecture will be streamed from these following links at Panopto:
Defence (12:15 - 15:00)
Trial Lecture (10:15 - 11:15)
The thesis is available Here