Brain Inspired Algorithms

Select By Year:

Select Another Subject Area:

Google Scholar

You can also find out about my published research on Google Scholar, which, as well as providing links to the papers and their abstracts, additionally provides the latest citation counts and index values.

Overview

After working for nearly a decade in sub-areas of artificial intelligence (robotics, temporal reasoning, constraint satisfaction and satisfiability), by 2006 I started to question the underlying idea that general artificial intelligence could arise by integrated all the sub-areas of AI into a unified Heath Robinson-type hybrid system. For the more neuroscience discovered about the one structure that we know exhibits general intelligence (i.e. the human neocortex) the less it looked like it implemented a hybrid set of independent specialised systems. Instead the neocortex showed itself to possess a remarkably uniform structure that seems to be performing the same basic function across its entire surface. And what has emerged from twenty-first century neuroscience is the idea that the neocortex is primarily engaged in making predictions about its own inputs and outputs – which has come to be known as the theory of predictive processing.

Taking heed of these developments, I formed the Cognitive Computing Unit at Griffith University as a centre of research where we could explore the computational implementation of predictive processsing and the following papers present the outcomes of this research. The primary focus has been the investigation and development of a particular implementation of predictive processing known as Hierarchical Temporal Memory, first proposed by Jeff Hawkins in the mid-2000s. This led to my own proposal of Hierarchical Temporal Intentionality (see ASSC 2014 below) which brings together the idea of predictive processing with the discoveries and insights of twentieth century phenomenology – connecting the brain science with our immediate experience of being conscious. This synthesis of science and consciousness is also a central theme in my new book The Questioning of Intelligence (see the Books page for more details, and The Being of Form below).

As with my work in constraint satisfaction, much of the actual coding and experimental evaluation reported in the papers below was performed by the PhD and Honours students I was supervising. And, as is usual, you can tell who had the greatest input into each paper by the order in which the author’s names are listed. Of these authors, I have been the primary PhD supervisor for Benjamin Cowley, Adam Kneller and Linda Main, and primary Honours supervisor for Andrew Srbic (for more detail on the pre-2010 papers co-authored with Jolon Faichney see the Handwriting Recognition page).

2021

Thornton, J. (2021). The Being of Form. In: The Questoning of Intelligence: A phenomenological exploration of what it means to be intelligent. FUBText: Brighton, pp. 320-350, ISBN 978-1-8384787-0-4

Excerpt: From this we can see that our scientific models of atomic being are not determined or limited by the dimensionalities of our perceptual systems. These dimensionalities only limit the extent to which we can picture or imagine what our scientific models ‘look like’. But if we consider the meaning of these models more fundamentally, we can see that they are still based on a quite specific conception of being—namely an idea of enduring mathematical form. This too is a human idea that is ‘built in’ to our perceptual systems. For, according to our latest scientific theories, these systems have evolved to detect invariance in the streams of sensorimotor input and output, in order to form predictions of the future evolution of these information streams. What emerges from this process are our ideas of a world populated with determinate and identifiable things and events. These are the entities we project to be the causes of the streams of sensorimotor information that continuously activate our nervous systems. Here, in order for something in the world to ‘be’ or to ‘exist’ it must possess the kind of enduring mathematical form that enables us to predict the shape of our future. This is all to do with computational efficiency. If you were to attempt to predict what you will experience in the next moment purely on the basis of the state of excitation of all your sensorimotor nerve impulses in this moment, it would turn out to be impossible. There is just not enough information in the previous state to determine with any accuracy what will happen in the next state. For virtually nothing at this level is stable. It is the river in which you cannot step twice. What the brain is doing is forming connections between its inputs and outputs in such a way as to detect the presence of enduring or stable relationships. Once it has encoded the essential underlying relationships that exist in the sensorimotor streams that emerge from our interactions with a particular object—such as the table in front of me now—then instead of trying to predict how the visual field is going to change as I shift my gaze over the table by examining the immediate state of that visual field, it uses what it has stored from past experience with tables in general, and with this table in particular, to predict and project how the table will appear in the next moment. Of course the immediate state of my visual system informs this process, but it is the mathematical form of the table that is encoded in the brain that determines what I am predicting, both in terms of what I expect to perceive and how I expect to act.

2018

Cowley, B., Thornton, J. R., Main, L. & Sattar, A. (2018). Precision without Precisions: Handling uncertainty with a single predictive model. In: ALIFE 2018, Proceedings of the 2018 Conference on Artificial Life, Tokyo, Japan, MIT Press, pp. 129-136.

2017

Cowley, B., Thornton, J. R., Main, L. & Sattar, A. (2017). Dynamic Thresholds for Self-Organizing Predictive Cells. In: ECAL 2017, Proceedings of the 14th European Conference on Artificial Life, Lyon, France, MIT Press, pp. 114-121.

Cowley, B. & Thornton, J. R. (2017). Feedback Modulated Attention within a Predictive Framework. ACALCI 2017, Artificial Life and Computational Intelligence, 3rd Australasian Conference. Lecture Notes in Computer Science, 10142, 61-73. Springer, ISSN 0302-9743.

Main, L. & Thornton, J. R. (2017). Stable Sparse Encoding for Predictive Processing. In: SSCI 2017, Proceedings of the 2017 IEEE Symposium Series on Computational Intelligence, Hawaii, USA, IEEE, pp. 1-8.

2015

Kneller, A. & Thornton, J. R. (2015). Distal Dendrite Feedback in Hierarchical Temporal Memory. In: IJCNN 2015, Proceedings of the 2015 International Joint Conference on Neural Networks, Killarney, Ireland, IEEE, pp. 385-392.

Main, L. & Thornton, J. R. (2015). A Cortically Inspired Model for Bioacoustics Recognition. ICONIP 2015: 22nd International Conference on Neural Information Processing, Istanbul, Turkey, Lecture Notes in Computer Science, 9492, 348-355, Springer, ISSN 0302-9743.

Thornton, J. R. (2015). The Transcendence of Computational Intelligence. PhD Confirmation Document, 141 pp. Completed under the supervision of John Mandalios, School of Humanities, Griffith University.

Abstract: The nature of the subject matter of this thesis makes it difficult to provide a concise summary of what I intend to write. My aim is rather to discover answers to the questions I am posing through the very process of writing. To begin therefore, I can only state the basic questions that are motivating my work: What is intelligence? What is computation? What is computational intelligence? What could it mean to transcend computational intelligence? Is such transcendence demonstrably possible? Of course I have presentiments of answers to these questions. Firstly, in agreement with Roger Penrose, I would say that any procedure that can be realised as a Turing machine can be considered computational. The question for the thesis is whether we can abstractly consider the functioning of the human brain to be equivalent to that of a (very sophisticated) Turing machine. To answer this I shall be examining the latest hierarchical predictive coding models of the functioning of the human neocortex and asking to what extent such (computational) models can account for the capacities of human intelligence. In particular, I shall be enquiring into our capacity to have a direct knowledge of what it is to be conscious, and asking whether such knowledge provides evidence of our being able to transcend the purely computationally determined actions of our neural systems. The confirmation document itself comprises an introductory chapter which provides an outline of the final thesis. At the moment I envisage a three-part manuscript. Part One is concerned with the problem of gaining explicit access to a state of pure, pre-reflective consciousness and thereby discovering what it means to have a direct knowledge of consciousness. In Part Two I plan to examine the computational foundations of intelligence in the functioning of the human neocortex, both from a neuroscientific and phenomenological perspective. Part Three is then intended to draw together Parts One and Two in a consideration of the possible transcendence of computational intelligence.
Following on from the introduction, Chapters 2-5 are all intended for inclusion in the final thesis. They cover Part One described above, with Chapter 2 providing a method of access, and Chapters 3-5 examining the work of Descartes, Schopenhauer and Husserl in relation to the material presented in Chapter 2. Finally, the appendices present two conference papers and one workshop paper that I have written and presented during my candidature. These papers cover various aspects of the questions I intend to address in Parts Two and Three, including the relevance of Heidegger to contemporary cognitive neuroscience (Appendix A), whether the activity of the human brain is causally closed under laws that determine the local low-level functioning of neural populations (Appendix B) and how contemporary models of neocortical functioning can be mapped onto Husserl’s phenomenological account of temporal consciousness (Appendix C).

2014

Cowley, B., Kneller, A. & Thornton, J. R. (2014). Cortically-Inspired Overcomplete Feature Learning for Colour Images. PRICAI 2014: Trends in Artificial Intelligence, 13th Pacific Rim International Conference on Artificial Intelligence. Lecture Notes in Computer Science, 8862, 720-732, Springer, ISSN 0302-9743.

Thornton, J. R. (2014). Hierarchical Temporal Intentionality. Abstract in: ASSC 18: Handbook of the 18th Conference of the Association for the Scientific Study of Consciousness, Brisbane, Australia, University of Queensland, pp. 41-42.

2013

Thornton, J. R. & Srbic, A. (2013). Spatial Pooling for Greyscale Images. International Journal of Machine Learning and Cybernetics. 4(3), 207-216, Springer, ISSN 1868-8071.

2012

Thornton, J. R., Main L. & Srbic, A. (2012). Fixed Frame Temporal Pooling. AI 2012: 25th Australasian Joint Conference on Artificial Intelligence, Lecture Notes in Computer Science, 7691, 707-718, Springer, ISSN 0302-9743.

2011

Thornton, J. R., Srbic, A., Main, L. & Chitsaz, M. (2011). Augmented Spatial Pooling. AI 2011: 24th Australasian Joint Conference on Artificial Intelligence, Lecture Notes in Computer Science. 7106, 261-270, Springer, ISSN 0302-9743.

2009

Thornton, J. R., Faichney, J., Blumenstein, M., Nguyen, V. & Hine, T. (2009). Offline Cursive Character Recognition: A state-of-the-art comparison. In: IGS 2009: Proceedings of the 14th Conference of the International Graphonomics Society, Dijon, France, pp. 148-151.

2008

Thornton, J. R., Faichney, J., Blumenstein, M. & Hine, T. (2008). Character Recognition using Hierarchical Vector Quantization and Temporal Pooling. AI 2008: 21st Australasian Joint Conference on Artificial Intelligence. Lecture Notes in Computer Science, 5360, 562-572, Springer, ISSN 0302-9743.

2006

Thornton, J. R., Gustafsson, T., Blumenstein, M. & Hine, T. (2006). Robust Character Recognition using a Hierarchical Bayesian Network. AI 2006: 19th Australian Joint Conference on Artificial Intelligence. Lecture Notes in Computer Science, 4304, 1259-1264, Springer, ISSN 0302-9743.