Unexpected Evolutionary Invariance: Understanding a Major Transition in the Evolution of Networks of Orientation Selective Neurons
by Joey (Zoe) Rowe Stawyskyj
Date of Examination:2024-11-28
Date of issue:2025-10-15
Advisor:Prof. Dr. Fred Wolf
Referee:Prof. Dr. Fred Wolf
Referee:Prof. Dr. Alexander Ecker
Persistent Address:
http://dx.doi.org/10.53846/goediss-11568
Files in this item
Name:Understanding_a_Major_Transition_in_Neural_C...pdf
Size:23.5Mb
Format:PDF
The following license files are associated with this item:
Abstract
English
The mechanisms of development and function of neural architectures involved in information detection and processing of sensory stimuli is still largely unknown, despite these processes being crucial in mediating interactions with the external world. While it is known that individual neurons are sensitive to, often very specific, external stimuli, a complete understanding of sensory processing requires understanding of the architecture created by the network of these neurons. These neural organisations can be understood through modelling with mathematical theories such as, where justified, optimisation theory. In mammalian primary visual cortex, form vision is dependent on networks of orientation selective neurons. In carnivores, primates, and marsupials, orientation domains comprised of local clusters of similarly tuned neurons are organised in quasi-periodic patterns called orientation preference maps. This architecture is quantitatively invariant among these highly separated lineages with substantial evolutionary time since their divergence. This unexpected invariance, which is re-examined and confirmed in this thesis, supports the use of optimality theories to assess development and function of these maps. Here is initially presented a series of manuscripts that contribute to the confirmation of this invariance. Firstly a method is presented for the analysis of intrinsic signal imaging data, the most common method of recording orientation preference maps, that allows for maximal comparability regardless of the species and origin of the data. Further, we provide methods for comparing the magnitude and nature of experimental noise. These methods are then applied to intrinsic signal measurement of a novel species, the fat-tailed dunnart (Sminthopsis crassicaudata), the second marsupial in which orientation preference maps have been investigated. This allows for greater understanding of the degree of spread of the invariant architecture, leading to insights about the development of the orientation preference maps. Quantitative assessment, accounting for low signal-to-noise ratios of the data, of the position and prevalence of pinwheel singularities in the maps suggest that dunnart orientation preference maps are of the same kind as those found in other mammalian species. This analysis, while the standard in the field, does not take into account other potential systematic map differences, including differences in the pattern of selectivity strength and differences in the size, shape and position of orientation clusters, all of which can vary without changing the pinwheel arrangement. Thus, a pixel-wise comparison method is defined in the third manuscript based on cosine similarity using patch sampling of these theoretically infinite maps and calculates a probability mass density of similar patches. The ensembles of probability mass densities can then be compared between different patterns using a rank-frequency test, accounting for the inevitable under-sampling. This confirmed invariance suggests that the architecture may be optimal for some computational or developmental goal. Two optimisation measures are proposed: minimisation of neural turnover during map development and spatial predictive coding, based on the ability of the network to represent information in distant areas about occluded regions of the visual field. Further a numerical method is proposed that is capable of large-scale, long time simulations with which these theories could be numerically confirmed. The spatial predictive information optimisation shows that some regions of parameter space have biologically realistic patterns, while it is shown using a mathematical pattern forming model that the biologically quantitatively invariant solution set optimal for minimal turnover. This suggests that minimal turnover is likely a relevant optimisation principle for orientation preference maps and that the spatial predictive coding may be one computationally optimality goal of the system. This likely explains features of the map, such as the heterogeneity of realisations mediated by a high dimensional ground state and the quasi-periodicity of the patterns. As such, these optimisation theories are capable of explaining both features of the patterns and the unexpected evolutionary invariance.
Keywords: Orientation preference maps (OPMs); Pattern formation; Pinwheels; Dynamical systems; Sensory neuroscience; Evolution of the visual system; Optimisation of dynamical systems; Evolutionary optimisation; Quasi-periodic patterns
