December, 13th – Marc Goodfellow
Title: Modelling pathological brain dynamics
Abstract:
Disorders of the brain can often result in alterations to its large-scale dynamics. An example is epilepsy, in which electrographic measurements display abnormal rhythms, particularly during seizures. Understanding why these dynamics are generated is challenging, particularly in the clinical setting, but better insight could help to improve diagnosis and treatment. In this talk I will discuss a particular approach to this problem, using mathematical models of large-scale brain networks to understand pathological dynamics. I will demonstrate how the study of such models can lead to new insight into the generation of seizures, and how models can be combined with clinical data to generate predictions for the surgical treatment of epilepsy.
December, 6th – Armin Lak
Title: Dopaminergic and prefrontal basis of learning from sensory confidence and reward value
Abstract:
Deciding between stimuli requires combining their learned value with one’s sensory confidence. We trained mice in a visual task that probes this combination. Mouse choices reflected not only present confidence and past rewards but also past confidence. Their behaviour conformed to a model that combines signal detection with reinforcement learning. In the model, the predicted value of the chosen option is the product of sensory confidence and learned value. We found precise correlates of this variable in the pre-outcome activity of midbrain dopamine neurons and of medial prefrontal cortical neurons. However, only the latter played a causal role: inactivating medial prefrontal cortex before outcome strengthened learning from the outcome. Dopamine neurons played a causal role only after outcome, when they encoded reward prediction errors graded by confidence, influencing subsequent choices. These results reveal neural signals that combine learned value with sensory confidence before choice outcome and guide subsequent learning.
November, 29nd – Nothing!!
November, 22nd – Bernhard Staresina
Title: Memory consolidation during sleep: Mechanisms and representations
Abstract:
In this talk, I will first present direct recordings from the human hippocampus during natural sleep. Analyses focus on the question how different sleep signatures (slow oscillations, spindles and ripples) interact and may facilitate hippocampal-neocortical information transfer. I will then turn to memory representations being reactivated during sleep. Using targeted memory reactivation, we show that sleep spindles seem to facilitate content-specific consolidation.
November, 15th – Jacques-Donald Tournier
Title: Multi-shell diffusion MRI and its applications in the neonatal brain
Abstract:
Recent advances in MRI acquisition now allow the routine acquisition of large amounts of so-called multi-shell diffusion MRI data within reasonable time frames. This opens up exciting new possibilities, but also brings additional challenges. This talk will present new methods for the acquisition and analysis of such data, both at the single-subject and at the group level. The talk will focus primarily (but not exclusively) on applications within the neonatal brain, using data acquired as part of the developing human brain connectome project.
November, 8th – Dan Goodman
Title: The Reluctant Machine Learner
Abstract:
The unique quality of the brain is that it can perform difficult tasks.
The traditional approach to modelling in neuroscience, though, has focussed on simple tasks, because those were the only ones we could model. Recently, that has all changed with the advent of powerful new methods from machine learning that can recognise some images better than humans, for example. I will argue that we have to study the brain solving difficult tasks, and therefore we have to be using techniques from machine learning because these are the only known methods that enable us to do that. However, that doesn’t mean that the brain is at all like the current best known machine learning models. Those models miss out on a lot of important points, like temporal dynamics and spiking neurons. Moreover, they make mistakes that humans would never make and require vastly more data than we do to learn. Despite these issues, neuroscience has a lot to gain from adopting machine learning methods, and I’ll talk about a couple of ongoing projects in my lab that attempt to use machine learning methods in a way that is more compatible with traditional neural modelling: modelling speech recognition in the auditory system; and trying to understand the computational role of the heterogeneity observed in real brains.
November, 1st – Christina Buetfuring
Title: Decision coding by layer 2/3 neurons in primary somatosensory cortex
Abstract:
Sensory information enables us to make informed choices that are critical for survival. While primary sensory areas provide information on sensory stimuli, behaviourally-relevant decision-making variables have been shown to be represented in higher-order association cortices. Therefore, sensory coding and decision-making are typically studied under the assumption of anatomical separation. Neurons in the superficial layers of the whisker region of primary somatosensory cortex (S1), barrel cortex, not only receive somatotopically mapped bottom-up inputs from the thalamorecipient layer 4 but also lateral projections from neighbouring barrels and top-down projections from higher cortical areas. Therefore, layer 2/3 (L2/3) neurons in barrel cortex are a prime candidate for providing an intersection of sensory processing and decision-making in complex behavioural tasks. Previous work using electrophysiological recordings in monkeys, rats and mice has not found conclusive choice activity in S1 but was limited to low number of neurons. Studies using two-photon calcium imaging found that some behavioural aspects modulate activity in L2/3 barrel cortex neurons. It is unclear, however, whether the signal difference across trial types in those studies reflects choice-related signals or a modulation of activity by action-related variables such as motivation, movement preparation etc. Here, we used two-photon calcium imaging of neurons in L2/3 mouse barrel cortex during a cued texture discrimination task with two lickports to determine whether these neurons can code for behaviourally-relevant decision variables. We found neurons carrying information about the stimulus irrespective of the behavioural outcome (‘stimulus neurons’) as well as neurons whose activity carried information about the choice to be made (‘decision neurons’). Choice-related activity in decision neurons is not driven by signals related to motor output, but instead follows stimulus presentation. Furthermore, ambiguous population coding of decision neurons predicts miss trials and an improvement in categorical coding in decision neurons coincides with learning the stimulus-choice association. Our identification of neurons encoding stimulus and behaviourally-relevant decision signals within the same circuit suggests a direct involvement of L2/3 S1 in the decision-making process.
Location: GEOG BLDG G.11N SR1
October, 25th – first year student projects
October, 18th – first year student projects
October, 11th – Cian O’Donnell
Title: Neural variability in Autism
Abstract:
Autistic people often have sensory processing deficits, and we would like to understand why. One clue comes from the observation that Autistic peoples’ EEG and fMRI responses to sensory stimuli are more variable than those in neurotypical people. We used in vivo two-photon calcium imaging of populations of layer 2/3 cortical neurons in young wild-type and Fragile-X Syndrome mouse models to search for three aspects of such variability at a cellular level: 1) across single trials from identical stimuli in the same animal, 2) across animals of the same age, and 3) longitudinally across days in the same animals. I will present what we found. Work with Beatriz Mizusaki (Univ of Bristol), Nazim Kourdougli, Anand Suresh, and Carlos Portera-Cailliau (Univ of California, Los Angeles).
Location: PHYS BLDG 3.34
October, 4th – Dimitris Pinotsis
Abstract:
In this talk, I will discuss how deep neural networks can reveal semantic and biophysical properties of memory representations in the brain (neural ensembles or cell assemblies).
First, I will consider a flexible decision-making paradigm and show that deep neural networks allow us to understand the sensory domains and semantics different brain areas prefer (motion vs color) and code (sensory signals vs abstract categories) respectively. These results will also suggest a way for studying sensory and categorical representations in the brain by combining behavioural and neural network models.
Then, I will show that deep neural networks can also reveal cortical connectivity in neural ensembles and explain a well-known behavioral effect in psychophysics, known as the oblique effect. This work will also introduce a new mathematical approach for identifying neural ensembles that exploits a combination of machine learning, biophysics and brain imaging.