Cambridge Memory Meeting 2016

Various cortical areas, most prominently the prefrontal cortex and the hippocampus, display diverse, stimulus-dependent neural responses during the delay period of tasks with a working memory component (Siegel et al. 2011). The timescales of these responses are much longer than single neuron timescales. However, the principles that determine how cortical circuits learn to make use of such representations are not fully understood. We assume that the brain learns an internal (generative) model of temporal sequences in its environment, such as the contingency between conditioned (CS) and un- conditioned stimuli (US) in trace conditioning. We construct and train a spiking neural network to perform inference and make predictions under such an internal model, with separate populations of neurons representing observed and latent variables. Neurons representing latent variables maintain a memory of a CS by triggering a traveling wave in the latent variable population, with a phase that is determined by the CS identity. The network learns to “surf the wave” and generate a response after the correct delay, by maximizing a lower bound on the likelihood of the model using variational inference. This allows a spiking network to learn long delays, more than a hundred times longer than single neuron timescales, and longer than in previous unsupervised learning models (Rezende and Gerstner, 2014; Brea et al. 2013).

Dale’s principle states that neurons in the nervous system have an exclusive physiological effect, each neuron either excites or inhibits all its synaptic targets. It is usually impossible for a neuron to excite some of its targets while inhibiting others. This principle has been known for more than fifty years, and yet there is no explanation for its function. Supported by theoretical analysis and experimental data, I propose that the function of Dale's principle is to maintain thermodynamic equilibrium in neural circuits. I study a nonlinear dynamical model characterized by a given synaptic matrix and input noise. Using the Fokker-Planck equation, I calculate the synaptic matrix consistent with thermodynamic equilibrium, and I show that it must satisfy Dale's principle under quite general assumptions. In order to test the theoretical predictions, I analyze the activity of neurons in the awake mammalian brain, and I show that collective neural dynamics displays temporal reversibility, which is a hallmark of thermodynamic equilibrium. I conclude by speculating on the significance of equilibrium and reversibility on brain function. While out-of-equilibrium dynamics may better support neural computation, equilibrium represents a desirable "idle" state.

Fear is a strong emotional experience. A memory of learned fear in rodents can be quantified using a Pavlovian fear-conditioning task. Once reactivated the fear memory can be expressed. However, the memory can also be destabilised and subsequently restabilised under certain conditions. Fear reminders engage the dopamine system in the basolateral amygdala (Yokoyama et al., 2005), but the contribution of dopamine signalling to the retrieval and destabilisation of fear memory is not fully understood. We explore the idea that dopamine is an essential modulator of fear memory through its ability to regulate synaptic mechanisms.
We perform a combination of behavioural testing and molecular analysis in rodents. The reactivation of a cued-fear memory activated the MAPK pathway and Extracellular Regulated Kinase (ERK) in the basolateral amygdala, and not it other brain regions. We analysed the regulation of this pathway, post memory reactivation, downstream of glutamate and dopamine receptor using pharmacological intervention directly into the amygdala to modulate these receptors.
The results will further our understanding of the molecular mechanisms downstream of dopamine receptor signalling for retrieval and destabilisation of memories.

Successful interactions in our environments entail extracting targets from cluttered scenes and discriminating highly similar objects. Previous functional MRI (fMRI) studies show differential activation patterns for learning to detect signal-in-noise vs. discriminate fine feature differences. However, fMRI does not allow us to discriminate excitatory from inhibitory contributions to learning. Recent Magnetic Resonance Spectroscopy (MRS) studies link γ-aminobutyric acid (GABA), the main inhibitory neurotransmitter, to learning processes and brain plasticity. Here we test whether BOLD changes correlate with GABA concentration changes after training in learning tasks. We trained observers to discriminate radial from concentric Glass patterns that were either presented in background noise or were highly similar to each other. Our results demonstrate that, for the signal-in-noise task, decreased GABA in occipito-temporal cortex correlates with increased BOLD changes and behavioural improvement after training. In contrast, for the fine discrimination task, increased GABA correlates with fMRI selective pattern activations and task performance. Our findings suggest dissociable visual learning mechanisms; that is, GABAergic inhibition may facilitate learning to discriminate fine differences by enhancing the tuning of feature selective neurons, while learning to see in clutter by suppressing background clutter and enhancing the gain in large neural populations.

(poster and talk - see abstract above)

Mixed selectivity is characterised by high-dimensional neural representations that are tuned to multiple task-related aspects. Neurons with such properties has been found in multiple regions in the brain and is hypothesised to support complex cognitive tasks. However, it is unclear what kind of computation mixed selectivity is optimal for and why it is optimal. Here we demonstrate that mixed selectivity is an intrinsic property of neural networks optimised for generative tasks through both analysis and simulations. By extending recently developed theories in deep learning, we interpret a generative neural network as a variational autoencoder, which can be decomposed into a recognition model that maps external input into representations and a generative model that reconstruct data from internal representations. This decomposition allows us to investigate neural representation, thus mixed selectivity, from a normative perspective. Our analytical results are supported by experiments on recurrent neural networks trained for sequence generation. In conclusion, we prove that mixed selectively is optimal for robustly generating data from neural networks, and demonstrated that it emerges spontaneously when optimising a neural network for generating data under noise.

Evidence accumulation and integration are the typical models for the use of time in perceptual decision-making. Here, we formulate two opposing probabilistic models of an orientation-estimation task in which stimulus time presentation is used for either evidence integration or probabilistic sampling. In evidence integration, time is used to iteratively collect evidence and update a posterior distribution over an environmental variable given the stimulus. Crucially, we assume that while the posterior distribution is changing as evidence is integrated it is represented exactly at each time point. Alternatively, under the sampling hypothesis time is used to approximately represent the posterior distribution by sequentially sampling from a static distribution. We devised a novel orientation-estimation task in which subjects also reported their uncertainty about their estimate. Therefore, on each trial we obtained a subjective measure of uncertainty along with a true measure of error. We developed intuitive as well as formal theoretical predictions of how humans should behave in this task given the two models and identified the across-trial correlation between error and uncertainty as a measure to distinguish between them. Finally, we collected data from multiple human subjects and obtained initial evidence supporting the sampling model.