Lech Szymanski, Our people, Department of Computer Science

Owheo Building, Room 249
Tel +64 3 479 5691
Email lech.szymanski@otago.ac.nz

My research interests are machine learning, deep representation and connectionist models. Our state of the art computation models can learn from examples, but not without the heavy involvement of an expert user, making critical decisions about system architecture, parameters, and appropriate representation. My ultimate objective is to make machine learning into a tool that is easier to use by an average user. However, in order for that to happen, we need to develop truly autonomous machine learning algorithms, capable of forming an appropriate model for the task at hand, completely on their own.

Before my PhD, which I completed in 2012, I worked as a software engineer for a wireless telecommunications company in Ottawa, Canada. My background is in computer and electrical engineering, with a focus on embedded programming and digital signal processing. My interest in artificial neural network models dates back to summer employment, while still an undergraduate student, developing programs for data analysis in a neurobiology laboratory at the National Research Council Canada. Since then I have worked on several aspects of modelling and machine learning, including speech recognition, classification, learning theory and object recognition from images.

For more information and selected publications, see this page

Publications

McAlister, H., Robins, A., & Szymanski, L. (2024). Prototype analysis in Hopfield networks with Hebbian learning. Neural Computation. Advance online publication. doi: 10.1162/neco_a_01704 Journal - Research Article

van der Vliet, W., Lal Khakpoor, F., Tetereva, A., Szymanski, L., & Pat, N. (2023). Bypassing parcellations when building predictive models for capturing cognitive abilities from task functional MRI. Proceedings of the Psycolloquy Symposium. (pp. 19). Dunedin, New Zealand: Department of Psychology, University of Otago. Retrieved from https://www.otago.ac.nz/psychology/research/otago059081.html Conference Contribution - Published proceedings: Abstract

van der Vliet, W. P., Lal Khakpoor, F., Tetereva, A., Szymanski, L., & Pat, N. (2023). Parcels or voxels? Better methods for predicting cognition from task-based fMRI. In K.-L. Horne (Ed.), Proceedings of the 39th International Australasian Winter Conference on Brain Research (AWCBR). (pp. 43). Retrieved from https://www.awcbr.org Conference Contribution - Published proceedings: Abstract

Xu, H., Szymanski, L., & McCane, B. (2023). VASE: Variational assorted surprise exploration for reinforcement learning. IEEE Transactions on Neural Networks & Learning Systems, 34(3), 1243-1252. doi: 10.1109/TNNLS.2021.3105140 Journal - Research Article

Szymanski, L., McCane, B., & Atkinson, C. (2022). Conceptual complexity of neural networks. Neurocomputing, 469, 52-64. doi: 10.1016/j.neucom.2021.10.063 Journal - Research Article

2024

Journal - Research Article

McAlister, H., Robins, A., & Szymanski, L. (2024). Prototype analysis in Hopfield networks with Hebbian learning. Neural Computation. Advance online publication. doi: 10.1162/neco_a_01704

2023

Journal - Research Article

Conference Contribution - Published proceedings: Abstract

2022

Journal - Research Article

Szymanski, L., McCane, B., & Atkinson, C. (2022). Conceptual complexity of neural networks. Neurocomputing, 469, 52-64. doi: 10.1016/j.neucom.2021.10.063

2021

Journal - Research Article

Atkinson, C., McCane, B., Szymanski, L., & Robins, A. (2021). Pseudo-rehearsal: Achieving deep reinforcement learning without catastrophic forgetting. Neurocomputing, 428, 291-307. doi: 10.1016/j.neucom.2020.11.050

Conference Contribution - Published proceedings: Full paper

Szymanski, L., & Lee, M. (2021). Coarse facial feature detection in sheep. Proceedings of the 36th International Conference on Image and Vision Computing New Zealand (IVCNZ). IEEE. doi: 10.1109/IVCNZ54163.2021.9653248

2020

Conference Contribution - Published proceedings: Full paper

van Sint Annaland, Y., Szymanski, L., & Mills, S. (2020). Predicting cherry quality using siamese networks. Proceedings of the 35th International Conference on Image and Vision Computing New Zealand (IVCNZ). IEEE. doi: 10.1109/IVCNZ51579.2020.9290674

Szymanski, L., & Lee, M. (2020). Deep sheep: Kinship assignment in livestock from facial images. Proceedings of the 35th International Conference on Image and Vision Computing New Zealand (IVCNZ). IEEE. doi: 10.1109/IVCNZ51579.2020.9290558

2019

Conference Contribution - Published proceedings: Full paper

Clark-Younger, H., Mills, S., & Szymanski, L. (2019). Stacked hourglass CNN for handwritten character location. Proceedings of the 2018 International Conference on Image and Vision Computing New Zealand (IVCNZ). IEEE. doi: 10.1109/IVCNZ.2018.8634694

Working Paper; Discussion Paper; Technical Report

Szymanski, L., McCane, B., & Atkinson, C. (2019). Switched linear projections and inactive state sensitivity for deep neural network interpretability. arXiv. Retrieved from https://arxiv.org/abs/1909.11275

Other Research Output

Szymanski, L. (2019, March). Folding, surprise and playing games: Deep learning at the CS department. Mathematics Seminar Series, Department of Mathematics & Statistics, University of Otago, Dunedin, New Zealand. [Department Seminar].

2018

Journal - Research Article

McCane, B., & Szymanski, L. (2018). Efficiency of deep networks for radially symmetric functions. Neurocomputing, 313, 119-124. doi: 10.1016/j.neucom.2018.06.003

Conference Contribution - Published proceedings: Full paper

Xu, H., McCane, B., & Szymanski, L. (2018). Twin bounded large margin distribution machine. In T. Mitrovic, B. Xue & X. Li (Eds.), Advances in artifical intelligence: Lecture notes in artificial intelligence (Vol. 11320). (pp. 718-729). Cham, Switzerland: Springer. doi: 10.1007/978-3-030-03991-2_64

Atkinson, C., McCane, B., & Szymanski, L. (2018). Increasing the accuracy of convolutional neural networks with progressive reinitialisation. Proceedings of the 2017 International Conference on Image and Vision Computing New Zealand (IVCNZ). IEEE. doi: 10.1109/IVCNZ.2017.8402457

Szymanski, L., & Mills, S. (2018). CNN for historic handwritten document search. Proceedings of the 2017 International Conference on Image and Vision Computing New Zealand (IVCNZ). IEEE. doi: 10.1109/IVCNZ.2017.8402461

Working Paper; Discussion Paper; Technical Report

Atkinson, C., McCane, B., Szymanski, L., & Robins, A. (2018). Pseudo-rehearsal: Achieving deep reinforcement learning without catastrophic forgetting. arXiv. 8p. Retrieved from https://arxiv.org/abs/1812.02464

Atkinson, C., McCane, B., Szymanski, L., & Robins, A. (2018). Pseudo-recursal: Solving the catastrophic forgetting problem in deep neural networks (v2). arXiv. Retrieved from https://arxiv.org/abs/1802.03875

2017

Conference Contribution - Published proceedings: Full paper

McCane, B., & Szymanski, L. (2017). Deep networks are efficient for circular manifolds. Proceedings of the 23rd International Conference on Pattern Recognition (ICPR). (pp. 3464-3469). IEEE. doi: 10.1109/ICPR.2016.7900170

Conference Contribution - Verbal presentation and other Conference outputs

Atkinson, C., McCane, B., & Szymanski, L. (2017, December). Increasing the accuracy of convolutional neural networks with progressive reinitialisation. Verbal presentation at the Electronics New Zealand Conference (ENZCon), Christchurch, New Zealand.

2016

Working Paper; Discussion Paper; Technical Report

Fu, X., McCane, B., Mills, S., Albert, M., & Szymanski, L. (2016). Auto-JacoBin: Auto-encoder Jacobian binary hashing. arXiv. 17p. Retrieved from https://arxiv.org/abs/1602.08127

2015

Journal - Research Article

Johnson, R., Szymanski, L., & Mills, S. (2015). Hierarchical structure from motion optical flow algorithms to harvest three-dimensional features from two-dimensional neuro-endoscopic images. Journal of Clinical Neuroscience, 22(2), 378-382. doi: 10.1016/j.jocn.2014.08.004

Conference Contribution - Verbal presentation and other Conference outputs

Knott, A., Szymanski, L., Gorman, C., & Takac, M. (2015, December). Predicative sentences and perceptual mechanisms. Verbal presentation at the Annual Conference of the Linguistic Society of New Zealand: Language & Society (LangSoc), Dunedin, New Zealand.

2014

Journal - Research Article

Szymanski, L., & McCane, B. (2014). Deep networks are effective encoders of periodicity. IEEE Transactions on Neural Networks & Learning Systems, 25(10), 1816-1827. doi: 10.1109/TNNLS.2013.2296046

Conference Contribution - Published proceedings: Full paper

Johnson, R., Mills, S., & Szymanski, L. (2014). Optical flow algorithms to recover 3D information from 2D endoscopic images. Proceedings of the 6th World Congress for Endoscopic Surgery of the Brain and Spine and Second Global Update on FESS, the Sinuses and the Nose (Endomilano). (pp. 76-77). Turin, Italy: Edizioni Minerva Medica. [Full Paper]

Mills, S., Szymanski, L., & Johnson, R. (2014). Hierarchical structure from motion from endoscopic video. Proceedings of the 29th International Conference on Image and Vision Computing New Zealand (IVCNZ). (pp. 102-107). New York: ACM. doi: 10.1145/2683405.2683411

Working Paper; Discussion Paper; Technical Report

Szymanski, L., & Eyers, D. (2014). Practical use of SELinux for enhancing the security of web applications [Technical Report OUCS-2014-02]. Dunedin, New Zealand: Department of Computer Science, University of Otago. 78p. Retrieved from http://www.cs.otago.ac.nz/research/techreports.php

2013

Conference Contribution - Published proceedings: Full paper

Szymanski, L., & McCane, B. (2013). Learning in deep architectures with folding transformations. Proceedings of the International Joint Conference on Neural Networks (IJCNN). IEEE. doi: 10.1109/IJCNN.2013.6706945

Martin, S., & Szymanski, L. (2013). Singularity resolution for dimension reduction. Proceedings of the 28th International Conference of Image and Vision Computing New Zealand (IVCNZ). (pp. 19-24). IEEE. doi: 10.1109/ivcnz.2013.6726986

2012

Conference Contribution - Published proceedings: Full paper

Szymanski, L., & McCane, B. (2012). Deep, super-narrow neural network is a universal classifier. Proceedings of the International Joint Conference on Neural Networks (IJCNN). IEEE. doi: 10.1109/IJCNN.2012.6252513

Szymanski, L., & McCane, B. (2012). Push-pull separability objective for supervised layer-wise training of neural networks. Proceedings of the International Joint Conference on Neural Networks (IJCNN). IEEE. doi: 10.1109/IJCNN.2012.6252366

Awarded Doctoral Degree

Szymanski, L. (2012). Deep architectures and classification by intermediary transformations (PhD). University of Otago, Dunedin, New Zealand. Retrieved from http://hdl.handle.net/10523/2129

2011

Conference Contribution - Published proceedings: Full paper

Szymanski, L., & McCane, B. (2011). Visualising kernel spaces. In P. Delmas, B. Wuensche & J. James (Eds.), Proceedings of the Image and Vision Computing New Zealand (IVCNZ) Conference. (pp. 449-452). Auckland, New Zealand: IVCNZ. [Full Paper]

2008

Conference Contribution - Published proceedings: Full paper

Szymanski, L., McCane, B., & Rountree, N. (2008). Maximum margin perceptron: Towards optimal and deterministic neural network architectures. Proceedings of the New Zealand Computer Science Research Student Conference. (pp. 266-269). [Full Paper]

2005

Conference Contribution - Published proceedings: Full paper

Szymanski, L., & Bouchard, M. (2005). Comb filter decomposition for robust ASR. Proceedings of the 6th Interspeech and 9th European Conference on Speech Communication and Technology (EUROSPEECH). (pp. 2645-2648). [Full Paper]

2001

Conference Contribution - Published proceedings: Full paper

Szymanski, L., & Yang, O. W. W. (2001). Spanning tree algorithm for spare network capacity. Proceedings of the Canadian Conference on Electrical and Computer Engineering. (pp. 447-452). IEEE. doi: 10.1109/CCECE.2001.933725