Emergence of Gaze-Invariant Spatial Representations and Coordinate Transformations in Spiking Neural Networks
The aim of this work is to investigate how a transformation from gaze-centred to head- or body-centred coordinates, and the invariance that comes with it, can emerge through local learning rules such as Spike-Timing-Dependent Plasticity (STDP), without any supervision. Visual input changes with every eye movement, even when an object stays in the same place in the world. A stable representation of space therefore requires combining retinal signals with information about eye position, a process thought to rely on gain modulation. You will study whether and how gain modulation, the coordinate transformation and gaze invariance can arise from the sensory and motor experience of a spiking neural network. The core of the thesis is to identify the factors that enable or influence this emergence. A central hypothesis concerns continuity. Objects do not disappear and reappear elsewhere: they follow trajectories that are continuous in space and time, whether they move slowly or quickly. You will investigate whether this continuity allows the network to discover that certain locations and signals are related, and how it contributes to learning. You will also examine the role of the eye movements themselves, including their statistics and whether they are self-generated, as well as the properties of the learning rules. You will evaluate the emergent representations on their invariance across gaze directions, their ability to distinguish locations, and their generalisation to new conditions. Finally, you will briefly consider how these representations can contribute to learning peri-personal space and the body schema.
A developing brain or an autonomous robot receives no supervisor specifying the correct coordinate transformation. Earlier models show that temporal association can produce head-centred responses, but which factors are necessary for this emergence, and in particular what role continuity plays, remain open questions for spiking networks. Answering them would inform theories of spatial development and could support robots that learn about their surroundings and their own bodies through experience.
Advisor:
- Pouya Abdollahzadeh Sadabad, pouya.abdollahzadeh1@…
Requirements:
- Strong proficiency in Python and experience with scientific simulation.
- Spiking Neural Networks: Basic understanding of neuron models and synaptic plasticity.
- Computational Neuroscience: Basic knowledge of visual processing and sensorimotor integration.
