Bayesian inference in spiking neural networks
The brain creates a coherent interpretation of the external world based on input from its sensory system. However, data from the senses are unreliable and confused. How does the brain synthesise its percepts?
Recent psychophysical experiments indicate that humans perform near-optimal Bayesian Inference (BI) in a wide variety of cognitive tasks, such as motion perception, decision making, and motor control (Körding & Wolpert, 2004; Kersten, Mamassian & Yuille, 2004; Knill & Richards, 1996. In BI, the likelihood of a particular interpretation of input signals from the senses is calculated, based not only on the sensory input signals, but also on the system's prior knowledge about the external world. If we could understand, in an engineering sense, how the brain accomplishes this, then we could apply this knowledge to the electronic sensors we build.
Neurons in the brain use action potentials (spikes) to communicate with each other. From calculations based on the energy consumption of the brain (only 20 watts in an adult human), it has been estimated that each cortical neuron fires only one spike per second, on average, while individual sensory neurons can fire close to 1000 spikes per second (Lennie, 2003). The question of how BI can be implemented using only spiking neurons with such slow communication rates is intriguing.
In the past five years a few papers have been published showing glimpses of how this could be achieved: Bobrowski, Meir & Eldar, 2009; Deneve, 2008a, 2008b; George & Hawkins, 2009, Ma, Beck, Latham & Pouget, 2006; Rao, 2004; Steimer, Maass & Douglas, 2009. If we were to take a similar approach to electronic sensor networks, we would minimise the bandwidth needed for communication and minimise power consumption.
At MARCS Institute for Brain, Behaviour and Development, our primary goal is to discover, through computational modelling, how the brain achieves near-optimal BI using networks of spiking neurons as the computational substrate. Two models will be investigated in parallel. First, we will develop a computational model in which each individual neuron performs an approximation of BI through the dynamic properties of its membrane potential and its synapses.
In this implementation, each neuron only needs to fire an action potential (spike) to inform other neurons that it has received evidence that alters the likelihood of the current interpretation of the sensory signals in a way that could not have been predicted from the previous output spikes of that neuron. If all inputs are consistent with the current interpretation, no spikes need to be sent. In other words, the neuron performs a form of predictive coding. This approach will lead to a system with very low spike rates.
We will also develop a second computational model of BI based on the novel approach of reservoir computing, in which the dynamical properties of networks of spiking neurons, rather than the properties of individual neurons, are used to implement the computations. We will combine this approach with that of the Hierarchical Temporal Memory approach, which has been shown to map to the known structure of cortical microcircuits in the neocortex accurately.
This approach is both more flexible and potentially more powerful than our first model, but needs more neurons and larger communication rates to implement BI. By developing both models, we will be able to choose which one is most suited for each particular smart sensor application.
From the software implementations of these two computational models, we will make testable predictions, both for sensory neurophysiology and psychophysics. For neurophysiology, these will take the form of membrane and synaptic time dynamics, while for psychophysics, these will take the form of specific human performance predictions on perceptual and decision tasks. One of the main advantages of the multidisciplinary nature of MARCS is that we will be able to test these predictions ourselves and use the results to improve our models. In turn, the model will guide psychophysical and neurophysiology research.
Delay adaptation in neural communication
The delay between the firing of a spike by one neuron to the reception of that spike by another neuron is typically in the range of one to forty milliseconds. These propagation delays have been largely ignored by both the neurophysiology and the neural computation communities. It has only very recently been recognised that the incorporation of delays enables a whole new class of computational systems, termed reservoir computing. Reservoir computing underlies quite possibly the computational architecture by which the brain performs Bayesian Inference.
To read out such reservoirs of neurons, propagation delays from neurons in the reservoir to the read-out neuron have to be precisely tuned. Some very limited evidence from neurophysiology for adaptive delays has begun to emerge, but in general delay adaptation in the brain has hardly been studied. The demonstration of a learning rule for adaptive delays in the brain would constitute a break-through discovery in neuroscience.
We investigate delay adaptation in vitro using rat brain slices. By patch clamping neurons at several points along their dendritic tree and their axons, we will be able to study if and how the spike propagation delay changes as a function of various input stimuli. We also use the neurophysiological data to build a computational model of delay adaptation to discover which learning rule applies and to discover the computational power of networks of neurons with adapting propagation delays.
Astrocytic modulation of brain waves
Brain waves are rhythmic voltage oscillations emerging from the synchronisation of individual neurons into a neuronal network. These oscillations emerge in all brain regions, and their patterns of synchrony and coherence underlie the neural code for sensory representation and short-term memory. Network oscillations range from slow to fast fluctuations, and are classified by power and frequency band, with different frequency bands being associated with specific behaviours, including attention, sleep and memory.
Cortical networks are constantly alternating between different dynamic states to accommodate the large rhythmic patterns underlying the diverse cognitive functions administrated by the cortex. However, the full extent of the functional structure of these networks, especially the interactions with astrocytic networks is poorly understood. It has been postulated that at least ten distinct mechanisms are required to cover the wide frequency range of neural oscillations, however, the mechanisms that gears the transition between distinct oscillatory frequencies are unknown.
In this study, we aim to discover the involvement of astrocytes, the prevailing subtype of glia in the brain, in modulating network activity. Specifically, we focus on astrocytic K+ clearance processes in modulating neural oscillations at both network and cellular levels. Since astrocytes are central for maintaining K+ homeostasis, our study suggests that modulation of their inherent capabilities to clear K+ from the extracellular milieu is a potential mechanism to optimise neural resonance behaviour and thus tune neural oscillations.
The impact of neuroinflammation on Alzheimer’s Disease
Alzheimer’s disease (A.D) is a neurodegenerative disorder characterized by significant impairment of cognitive function, memory loss and behavioural phenotypes such as anxiety and depression. There are several hypotheses regarding the etiology of A.D. The first and oldest hypothesis is the "cholinergic hypothesis" which suggest that a serious loss of cholinergic function in the basal forebrain and the associated loss of cholinergic innervation in the hippocampus and neocortex contribute significantly to the cognitive symptoms associated with AD. However, the source underling the loss of cholinergic cells is still unknown. Recent findings suggest that neuroinflammation is a preliminary process, which play a role in the onset of Alzheimer’s disease. However, the impact of neuroinflammation on cholinergic neurons is still an undiscovered area. In this project we aim to determine the effects of chronic and acute inflammatory processes on neurophysiological properties of the basal forebrain cholinergic system, and the susceptibility of cholinergic neurons during aging.
Targeting excitability of motoneuron’s in ALS mice models
Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease in which patient’s lose motor functions due to progressive loss of motor neurons in the brainstem and spinal cord. Evidence from patients in the clinic suggests that prior to the presentation of clinical symptoms, familial and sporadic ALS patients display an increase in neuronal hyperexcitability, however, the factors that instigate the changes in neural conductivity over the course of disease onset and progression are not well understood. In this project, we aim to investigate the ionic mechanisms governing changes in neuronal excitability in motor neurons of mice model for ALS. Specifically, we will assess the connectivity and function of astrocytes in the vicinity of these motor neurons and identify their effect on neuronal excitability in ALS.