Marthinus Theunissen

Timeline

1. 2020

   AAAI paper on cooperating classifiers

   As a PhD student at North-West University, they travel to New York in 2020 to present "DNNs as Layers of Cooperating Classifiers" at AAAI-2020. The paper, coauthored with Marelie Davel, Etienne Barnard and Arnold Pretorius, shows how fully connected feedforward networks can be viewed as the combination of one continuous and one discrete information processing system.

2. 2020

   Benign interpolation of noise

   In 2020, with Marelie Davel and Etienne Barnard, they publish "Benign interpolation of noise in deep learning" in the South African Computer Journal. Experiments with overparameterised multilayer perceptrons and controlled training data noise show that models are optimised for training data modularly and tend to fit uncorrupted samples first.

3. 2021

   A PhD on generalization

   In 2021, they complete a PhD in Computer and Electronic Engineering at North-West University in South Africa, supervised by Etienne Barnard and Marelie Davel. The thesis, "Generalization in deep learning: bilateral synergies in MLP learning", examines how feed-forward networks generalize to out-of-sample data.

4. February 2023

   The Missing Margin study

   In February 2023, they post "The Missing Margin: How Sample Corruption Affects Distance to the Boundary in ANNs" to arXiv with Coenraad Mouton and Marelie Davel. The empirical study shows that some types of training samples are modelled with consistently small margins while affecting generalization in different ways, using noise-corrupted MNIST and CIFAR10 data.

5. 2024

   Constrained input margins

   In 2024, with Coenraad Mouton and Marelie Davel, they publish "Input Margins Can Predict Generalization Too" in the Proceedings of the AAAI Conference on Artificial Intelligence. The paper defines "constrained margins", an input-margin measure that achieves highly competitive scores on the PGDL benchmark and outperforms other margin measurements in general.

6. 2024

   A fragmentation complexity measure

   Later in 2024, they coauthor "Is network fragmentation a useful complexity measure?" for the NeurIPS 2024 Workshop SciForDL, with five coauthors including Coenraad Mouton and Marelie Davel. The paper studies fragmentation, where a deep network's model function rapidly changes class across the input space, and shows that a fragmentation-based complexity measure achieves good performance on the PGDL benchmark.

7. July 2025

   The KnowIt toolkit

   In July 2025, they publish "KnowIt: Deep Time Series Modeling and Interpretation" on arXiv with R. Rabe, H.L. Potgieter and Marelie Davel. KnowIt is a Python toolkit for building deep time series models and interpreting them, and it separates the dataset, network architecture and interpretability technique through defined interfaces.

8. As of October 2026

   Senior lecturer at North-West University

   In June 2026, they coauthor "Data Pruning: Redundant, Problematic, and Interdependent Samples" with Leon Freese, which shows that two popular data pruning methods fail in the presence of significant label noise. They are a senior lecturer at North-West University, listed there as Dr Tian Theunissen, and work on generalization in deep learning and the interpretability of deep time series models.