Steve Kroon

Bio

Bio last updated: July 2022

Please note: The information below is unfortunately dated. I will try to update when I get time.

Curriculum Vitae (last updated: 25 April 2019)

Research

Google Scholar profile

Main research topics

• Generative modeling
• Bayesian methods
• Neural networks
• Computational intelligence in games

Potential students

Research output

• Making Superhuman AI More Human in Chess (with Daniel Barrish and Brink van der Merwe), Advances in Computer Games (ACG) 2023 (held online). Part of the Lecture Notes in Computer Science book series (LNCS, volume 14528). Work on creating chess agents that more faithfully imitate strong human players.
• Topological Dynamics of Functional Neural Network Graphs During Reinforcement Learning (with Martin Muller and Stephan Chalup), International Conference on Neural Information Processing (ICONIP) 2023. An investigation of topological structures present in graphs derived from the activation patterns in the neural networks obtained when training reinforcement learning agents.
• Integrating Bayesian Network Structure into Residual Flows and Variational Autoencoders (with Jacobie Mouton). Transactions on Machine Learning Research. Combines Jacobie's two 2022 ICLR workshop papers with some additional work from her Master's thesis.
• SIReN-VAE: Leveraging Flows and Amortized Inference for Bayesian Networks (with Jacobie Mouton). 2022 ICLR Workshop on Deep Generative Models for Highly Structured Data. Jacobie's poster on this work also won a book prize at the 2022 Deep Learning Indaba. (ICLR workshop poster| Indaba poster)
• Graphical Residual Flows (with Jacobie Mouton). 2022 ICLR Workshop on Deep Generative Models for Highly Structured Data. (Poster)
• SplyCI: Integrating Spreadsheets by Recognising and Solving Layout Constraints (with Dirko Coetsee, McElory Hoffmann, and Luc de Raedt), 19th Symposium on Intelligent Data Analysis (IDA 2021), Porto, Portugal. Part of the Lecture Notes in Computer Science book series (LNCS, volume 12695); also part of the Information Systems and Applications, incl. Internet/Web, and HCI book sub series (LNISA, volume 12695). (Download link from KU Leuven.) Also accepted for oral and poster presentation at the 2021 ECMLPKDD Workshop on Automating Data Science (ADS2021). SplyCI is an early prototype for automatic merging of spreadsheets; Dirko wrote this blog post on the article.
• If Dropout Limits Trainable Depth, Does Critical Initialisation Still Matter? A Large-scale Statistical Analysis on ReLU Networks (with Arnu Pretorius, Elanvan Biljon, Benjamin van Niekerk, Ryan Eloff, Matthew Reynard, Steve James, Benjamin Rosman, and Herman Kamper), Pattern Recognition Letters, Volume 138, October 2020, pages 95--105. (Elsevier Sharelink valid until 17 September 2020.)) Signal propagation theory establishes a limit on training depth for ReLU networks in the presence of noise regularization. Here, we employ an experimental design inspired by randomized controlled trials; our empirical analysis compares the critical initialization strategy to other strategies that are viable given the depth limit resulting from noise regularization.
• On the Expected Behaviour of Noise Regularized Deep Neural Networks as Gaussian Processes (with Arnu Pretorius and Herman Kamper), Pattern Recognition Letters, Volume 138, October 2020, pages 75--81. (Elsevier Sharelink valid until end of August 2020.) This paper considers the impact of noise regularization (e.g. dropout) on the neural network Gaussian processes arising as the infinite-width limit of deep neural networks.
• Performance-Agnostic Fusion of Probabilistic Classifier Outputs (with Jordan Masakuna and Simukai Utete), 23rd International Conference on Information Fusion (FUSION 2020). This article presents a novel approach to once-off fusion of disparate classifier predictions to obtain a single consensus prediction. (Code repository)
• Stochastic Gradient Annealed Importance Sampling for Efficient Online Marginal Likelihood Estimation (with Scott Cameron and Hans Eggers), Entropy, Volume 21(11) (2019). (This was a special issue for extended versions of selected papers at MaxEnt 2019.) Proposes a new evidence estimation technique which can take advantage of mini-batching. This yields an online evidence estimation technique which is considerably faster than nested sampling and regular annealed importance sampling. (Version on arxiv.) Update (24 July 2020): Post-publication communication with Sam Power revealed that part of the approach used in this paper (and the original MaxEnt paper) corresponds closely to Nicolas Chopin's iterated batch importance sampling technique from 2002.
• Stabilising priors for robust Bayesian deep learning (with Felix McGregor, Arnu Pretorius and Johan du Preez). Extended abstract accepted for poster presentation at the NeurIPS 2019 Bayesian Deep Learning workshop in Vancouver, Canada.
• A Sequential Marginal Likelihood Approximation using Stochastic Gradients (with Scott Cameron and Hans Eggers), Proceedings, Volume 33(1) (2019). Paper accepted for poster presentation at MaxEnt 2019 in Garching, Germany, where it won second prize in the poster competition. Proposes a new evidence estimation technique based on a sequential decomposition of data, and using stochastic gradient MCMC approaches to estimate the factors required. This idea was extended and improved in "Stochastic Gradient Annealed Importance Sampling for Efficient Online Marginal Likelihood Estimation".
• Dropout Initialization (with Arnu Pretorius, Elan van Biljon, Ryan Eloff, Matthew Rynard, Benjamin van Niekerk, Steven James, Benjamin Rosman, and Herman Kamper). Poster presentation by Elan van Biljon at 2019 Deep Learning IndabaX South Africa, Durban
• A Coordinated Search Strategy for Solitary Robots (with Jordan Masakuna and Simukai Utete), Third IEEE International Conference on Robotic Computing (IRC), 2019, Naples. This paper proposes a strategy based on cellular decomposition of an unknown region for coordinating solitary robots, i.e. robots that mostly function alone, but interact with others when serendipitous encounters arise. The aim is for the strategy to reduce redundant search to enhance the search effectiveness of a group of such solitary robots. An extended version with more detail is available on the arxiv. (Code repository)
• Critical Initialisation for Deep Signal Propagation in Noisy Rectifier Neural Networks (with Arnu Pretorius, Elan van Biljon and Herman Kamper), Thirty-second Conference on Neural Information Processing Systems (NIPS), 2018, Montreal. (Accepted for poster presentation.) Poster based on preliminary version of this work, "Deep signal propagation for noisy rectifier neural networks", also presented by Arnu and Elan at the 2nd Deep Learning Indaba 2018 (Arnu was awarded US$1000 in Google Compute credits, while Elan was awarded an NVidia Titan V GPU for this poster). Poster also presented at Data, Learning and Inference (DALI), 2019, George, as well as the 2019 Machine Learning Summer School in Stellenbosch.
• Variational Autoencoders for Missing Data Imputation with Application to a Simulated Milling Circuit (with John McCoy and Lidia Auret), 5th IFAC Workshop on Mining, Mineral and Metal Processing, 2018, Shanghai. IFAC-Papersonline, Vol. 51 (21), pp. 141-146, 2018. (DOI: 10.1016/j.ifacol.2018.09.406). Poster based on this work presented by John at the 2nd Deep Learning Indaba 2018 (awarded US$1000 in Google Compute credits for the poster).
• Learning Dynamics of Linear Denoising Autoencoders (with Arnu Pretorius and Herman Kamper), 2018 International Conference on Machine Learning, 2018, Stockholm. (Oral presentation and poster. Link to paper and supplementary material on PMLR.) Preliminary work towards this presented as a poster, "Learning dynamics of regularised linear neural networks", by Arnu at the 1st Deep Learning Indaba 2017 - this poster won a book prize.
• No evidence for extensions to the standard cosmological model (with Alan Heavens, Yabebal Fantaye, Elena Sellentin, Hans Eggers, Zafiirah Hosenie, and Arrykrishna Mootoovaloo), Physical review letters, Volume 119, Number 10, pages 101301-1--101301-5 (2017). This article applies a proposed approach to computing Bayesian evidence to a variety of cosmological models. The approach is able to reuse the MCMCs used for parameter inference, and finds the standard cosmological model to be favoured over all other alternatives.
• New Reinforcement Learning Algorithm for Robot Soccer (with Moonyoung Yoon and James Bekker), ORiON, Volume 13, Number 1, Pages 1-20 (2017) (DOI: 10.5784/33-1-542). This article proposes a new reinforcement learning algorithm called temporal difference value iteration with state-value functions, and presents applications to decision-making problems in the RoboCup Small Size League.
• Unsupervised Pre-training for Fully Convolutional Neural Networks (with Stiaan Wiehman and Hennie de Villiers), 2016 PRASA-Robmech International Conference, 2016, Stellenbosch.
• Combining Tree Kernels and Word Embeddings for Plagiarism Detection (with Niël Thom and Brink van der Merwe), 2016 PRASA-Robmech International Conference, 2016, Stellenbosch. Submitted to work in progress stream, accepted for poster presentation.
• Tackling Inconsistency in Classifier Fusion (with Jordan Masakuna and Simukai Utete), Southern Africa Mathematical Sciences Association Annual Conference (SAMSA 2016), 2016, Pretoria. Poster presentation. Poster on this work also presented by Jordan Masakuna at 1st Deep Learning Indaba 2017 - this poster won a book prize.
• DSaaS: A Cloud Service for Persistent Data Structures (with Pierre le Roux and Willem Bester), CLOSER, 2016, Rome. This paper presents DSaaS, a web service providing confluently persistent data structures.
• A Comparison of Low-Cost Monocular Vision Techniques for Pothole Distance Estimation (with Sonja Nienaber and Thinus Booysen), 2015 IEEE Symposium Series on Computational Intelligence, December 2015, Cape Town (DOI: 10.1109/SSCI.2015.69). This article empirically compares the accuracy of some simple approaches to estimating the depth of objects in a ground plane from single images, with the application being estimating the distance to potholes. This article also introduced the annotated pothole data set used in Sonja's thesis (DOI: 10.13140/RG.2.1.3646.1520). This data has since been used as the basis for various hackathons and the like, including: an MIIA Hackathon in September 2017 (data), a Zindi challenge designed for Lederex and AI Expo in September 2019 (and subsequently kept open to the public and covered in a blog post by Zindi's johno), and two Kaggle datasets.
• N-Gram Representations For Comment Filtering (with Dirk Brand, Brink van der Merwe, and Loek Cleophas), 2015 Conference of the South African Institute for Computer Scientists and Information Technologists, September 2015, Stellenbosch (DOI: 10.1145/2815782.2815789). This article investigates the performance of N-gram models for predicting comment quality on a news website.
• Detecting Potholes Using Simple Image Processing Techniques and Real-world Footage (with Sonja Nienaber and Thinus Booysen), 2015 South African Transport Conference, 2015, Pretoria. This articles gives an image processing pipeline for detecting potholes from camera images under good visibility conditions. South African provisional patent 2015/04540 assigned to MTN in 2015 with authors as inventors, titled "A DEVICE AND METHOD OF DETECTING POTHOLES".
• Sample Evaluation for Action Selection in Monte Carlo Tree Search (with Dirk Brand), 2014 Conference of the South African Institute for Computer Scientists and Information Technologists, September 2014, Pretoria, pages 314-322. This article proposes leveraging a (possibly poor) evaluation function for selecting nodes to expand in MCTS when the branching factor is very high by selecting a node with the best evaluation from a sample. The proposal is evaluated on the board Game Risk.
• Decision Trees for Computer Go Features (with Francois van Niekerk), 2013 International Joint Conference on Artificial Intelligence Workshop on Computer Games, August 3, Beijing. This article proposes using decision trees for extracting domain knowledge in the context of Monte Carlo Tree Search for Computer Go.
• Binary Jumbled String Matching for Highly Run-Length Compressible Texts (with Golnaz Badkobeh, Gabriele Fici, and Zsuzsanna Lipták), Information Processing Letters, Volume 113, Issue 17, 30 August 2013, Pages 604–608 (DOI: 10.1016/j.ipl.2013.05.007). This article presents an alternative algorithm for jumbled pattern matching on binary strings making use of a different index to previous approaches. Typically, the index is smaller than the traditional approach, but query times are no longer constant. The advantage of our approach is most pronounced for strings with short run-length encoding.
• Monte-Carlo Tree Search Parallelisation for Computer Go (with Francois van Niekerk, Gert-Jan van Rooyen, and Cornelia Inggs), Proceedings of the 2012 Annual Research Conference of the South African Institute for Computer Scientists and Information Technologists, October 2012. This article presents results of parallelisation of Monte-Carlo Tree Search for multi-core and cluster systems in the Computer Go program Oakfoam. (Some results from this paper were also presented by Francois van Niekerk at a talk at the 2012 International Go Symposium, "New Work on MCTS Parallelisation and The State of the Art of Supercomputer Go and its Future".)
• Unsupervised Construction of Topic-based Twitter Lists (with Francois de Villiers and McElory Hoffmann), Proceedings of the 2012 ASE/IEEE International Conference on Social Computing (SocialCom 2012), Amsterdam, Netherlands, September 2012. This article investigates compares different document representation techniques, similarity measures, and clustering methods for unsupervised list construction according to topics tweeted about in Twitter.
• A Community-Based Model of Online Social Networks (with Leendert Botha), Proceedings of the 4th SNA-KDD Workshop on Social Network Mining and Analysis (SNAKDD 2010), Washington D.C., USA, July 2010 (accepted). (Also a poster presentation at the Eighth Workshop on Mining and Learning with Graphs (MLG-2010), Washington D.C., USA, July 2010.) This article describes a model for generating social networks with similar properties to existing social networks.
• Generalizing the Margin Concept to Arbitrary Classifiers (with Sarel Steel), Proceedings of the 57th Session of the International Statistics Institute, Durban, South Africa, August 2009 (abstract) - This paper highlights a result from my Ph.D. thesis, which extends the margin concept from thresholding real-valued classifiers to classifiers in general metric spaces.
• A PAC-Bayesian Generalization of a Result of Devroye (with Sarel Steel), Slides from a talk presented at the South African Statistical Conference, 2008 - This presentation highlights selected results from my PhD thesis concerning classical covering number bounds: a classical covering number bound for classification is generalized to regression and arbitrary ghost sample sizes, inter alia, and it is shown the resulting bound is actually a form of PAC-Bayesian bound employing a transductive "prior".
• A Framework for Estimating Risk, Ph.D. Thesis, Stellenbosch University, 2008 - This research was related primarily to risk estimation, notably constructing confidence intervals for the risk of a fitted model from the data used to fit the model.
• Getting to Grips with Support Vector Machines: Theory (with Christian Omlin), South African Statistical Journal, 38(2), pp. 93-114, 2004 - An introduction to the theory underlying SVMs, based on material in my Masters thesis.
• Getting to Grips with Support Vector Machines: Application (with Christian Omlin), South African Statistical Journal, 38(2), pp. 159-172, 2004 - Illustrates the use of LIBSVM for analyzing a simple data set.
• Support Vector Machines, Generalization Bounds and Transduction, Masters Thesis, Stellenbosch University, 2003 - Surveys the theory of Support Vector Machines and Generalization Bounds, and presents a transductive generalization bound.
• Putting the SVM in context (with Sarel Steel), Slides from a talk presented at the South African Statistical Conference, 2003 - This presentation presents the SVM from the regularization viewpoint, contrasting it with other well-known regularization techniques such as the lasso.
• Bounding Generalization of Support Vector Machines (with Christian Omlin), Slides from a talk presented at the South African Statistical Conference, 2003 - This presents the core ideas behind classical covering number bounds, applied to SVMs.
• A Covering Number PAC Bound for Transductive Problems with Applications to SVMs (with Christian Omlin), Unpublished manuscript, 2003 - This contains the transductive bound for SVMs presented in my Masters thesis.

Talks at colloquiua, seminars, etc.

• Decision Trees for Computer Go Features, A presentation based on joint work with Francois van Niekerk at University of Maastricht's Department of Knowledge Engineering, June 2014.
• The Bias-Variance Dilemma and Regularization Paths, Slides from a series of 3 seminars presented to Vision and Learning at Stellenbosch research group, October-November 2009 - Covers bias-variance decomposition of mean-squared error, the role of regularization in this context, and then investigates the relationships between forward stagewise modelling, least angle regression, and the lasso.
• Margin Bounds for Arbitrary Classifiers, Slides from a seminar presented to Vision and Learning at Stellenbosch research group, September 2009 - This presentation was a slight revision of my ISI presentation earlier in the year.
• Covering Number Bounds and Statistical Learning Theory, Slides from a seminar presented to Vision and Learning at Stellenbosch research group, July 2009 - This presentation was an introduction to core concepts in deriving covering number bounds and their relationship to basic ideas of uniform laws of large numbers in statistical learning theory.
• Bounding Generalization of Support Vector Machines, Slides from Department of Statistics Seminar series, Stellenbosch University, 2003 - This presents the core ideas behind classical covering number bounds, applied to SVMs.
• An introduction to Support Vector Machines, Slides for a Masters course in Intelligent Systems at the Department of Computer Science, Stellenbosch University, 2000 - Very simple introduction to SVMs

Graduate students

Project notes for honours students w.r.t. supervision (largely applicable to Masters and PhD students as well). (Last updated: 25 October 2012.)

Current Masters students

• Sonja Nienaber (from 2014): Sonja is working on pothole detection from a windscreen-mounted camera.
• Hilgard Bell (from 2012): Hilgard is working on using Monte-Carlo Tree Search trees to assess difficulty of puzzles for procedural content generation.

Completed Masters students

• Moonyoung Yoon (April 2015): Moonyoung worked on using reinforcement learning to acquire skills in robotic soccer.
• Francois van Niekerk (April 2014): Francois worked on decision trees for computer Go features. Now at Clockwork Acorn. (Thesis notes: the q subscripts for w, l, d, and r in Section 3.4.1 are unnecessary, since they are the same for all candidate queries; in description of the NEW query for stone graphs, "edge" should be "side".)
• Francois de Villiers (April 2013): Francois worked on recommender systems with social media.
• Leendert Botha (December 2011): Modeling Online Social Networks using Quasi-clique Communities. Now at Google.
• Andre Kriek (March 2009): RoboCup Formation Modeling. Now at 4i Software Development.

Old pages

• Course on Probability and Computing (2nd semester 2009)
• Course on Artificial Intelligence (1st semester 2006)

This page hosted on bigmetal.