Search

We present a joint deep neural system identification model for two major sources of neural variability: stimulus-driven and stimulus-conditioned fluctuations. To this end, we combine (1) state-of-the-art deep networks for stimulus-driven activity and (2) a flexible, normalizing flow-based generative model to capture the stimulus-conditioned variability including noise correlations. This allows us to train the model end-to-end without the need for sophisticated probabilistic approximations associated with many latent state models for stimulus-conditioned fluctuations. We train the model on the responses of thousands of neurons from multiple areas of the mouse visual cortex to natural images. We show that our model outperforms previous state-of-the-art models in predicting the distribution of neural population responses to novel stimuli, including shared stimulus-conditioned variability. Furthermore, it successfully learns known latent factors of the population responses that are related to behavioral variables such as pupil dilation, and other factors that vary systematically with brain area or retinotopic location. Overall, our model accurately accounts for two critical sources of neural variability while avoiding several complexities associated with many existing latent state models. It thus provides a useful tool for uncovering the interplay between different factors that contribute to variability in neural activity.

Deep neural networks (DNN) have set new standards at predicting responses of neural populations to visual input. Most such DNNs consist of a convolutional network (core) shared across all neurons which learns a representation of neural computation in visual cortex and a neuron-specific readout that linearly combines the relevant features in this representation. The goal of this paper is to test whether such a representation is indeed generally characteristic for visual cortex, i.e. gener-alizes between animals of a species, and what factors contribute to obtaining such a generalizing core. To push all non-linear computations into the core where the generalizing cortical features should be learned, we devise a novel readout that reduces the number of parameters per neuron in the readout by up to two orders of magnitude compared to the previous state-of-the-art. It does so by taking advantage of retinotopy and learns a Gaussian distribution over the neuron’s receptive field po-sition. With this new readout we train our network on neural responses from mouse primary visual cortex (V1) and obtain a gain in performance of 7% compared to the previous state-of-the-art network. We then investigate whether the convolutional core indeed captures general cortical features by using the core in transfer learning to a different animal. When transferring a core trained on thousands of neurons from various animals and scans we exceed the performance of training directly on that animal by 12%, and outperform a commonly used VGG16 core pre-trained on imagenet by 33%. In addition, transfer learning with our data-driven core is more data-efficient than direct training, achieving the same performance with only 40% of the data. Our model with its novel readout thus sets a new state-of-the-art for neural response prediction in mouse visual cortex from natural images, generalizes between animals, and captures better characteristic cortical features than current task-driven pre-training approaches such as VGG16.

Computational human head models have been used in studies of brain stimulation. These models have been able to provide useful information that can’t be acquired or difficult to acquire from experimental or imaging studies. However, most of these models are purely volume conductor models that overlooked the electric excitability of axons in the white matter of the brain. We hereby combined a finite element (FE) model of electroconvulsive therapy (ECT) with a whole-brain tractography analysis as well as the cable theory of neuronal excitation. We have reconstructed a whole-brain tractogram with 2000 neural fibres from diffusion-weighted magnetic resonance scans and extracted the information on electrical potential from the FE ECT model of the same head. Two different electrode placements and three different white matter conductivity settings were simulated and compared. We calculated the electric field and second spatial derivatives of the electrical potential along the fibre direction, which describes the activating function for homogenous axons, and investigated sensitive regions of white matter activation. Models with anisotropic white matter conductivity yielded the most distinctive electric field and activating function distribution. Activation was most likely to appear in regions between the electrodes where the electric potential gradient is most pronounced.

We consider the problem of learning a density function from observations of an unknown underlying model in a distributed setting, where the observations are partitioned into different sites. Applying commonly used density estimation methods such as Gaussian Mixture Model (GMM) or Kernel Density Estimation (KDE) to distributed data leads to an extensive amount of communication. A familiar approach to address this issue is to sample a small subset of data and collect them into a central node to run the density estimation algorithms on them. In this paper, we follow an alternative to the sub-sampling approach by proposing the nested Log-Poly model. This model provides an accurate density estimation from a small sized statistic of the entire data. In distributed settings, it transfers the small sized statistics from the client nodes to a central node. The estimation process is then run in the central node. The proposed model can be used in different learning tasks such as classification in supervised learning and clustering in unsupervised learning. However, the properties of nested Log-Poly make it a suitable model for one-dimensional density estimations in the distributed settings. This makes Log-Poly a good choice for naive Bayes classifier, where one-dimensional density estimation is required for every feature conditioned on the class label. We provide a theoretical analysis of the efficiency of our model in estimating a wide range of probability density functions. Our experiments show that nested Log-Poly outperforms the state of the art density estimators on several synthetic datasets. We compare the accuracy and the communication load of naive Bayes classifier using nested Log-Poly and other related density estimators on several real datasets. The experimental outcomes depict that nested Log-Poly has less communication load, while maintaining a competitive classification accuracy compared to similar methods that use the entire data. Moreover, we present a comprehensive comparison between nested Log-Poly and validated KDE with sub-sampling, in terms of the number of communicated variables and the number of bytes transferred between the clients and the central node. Nested Log-Poly provides comparable accuracy with the validated KDE with sub-sampling, while communicating fewer variables. However, our method needs to compute and transmit the variables with a high precision in order to accurately capture the details of the underlying distributions.

Computational human head models have been used in electrophysiological studies, and they have been able to provide useful information that is unable or difficult to acquire from experimental or imaging studies. However, most of these models are purely volume conductor models that overlooked the electric excitability of axons in the white matter of the brain. This study combined a finite element (FE) model of electroconvulsive therapy (ECT) with a whole-brain tractography analysis as well as the cable theory of neuronal excitation. We have reconstructed a whole-brain tractogram with 500 neural fibres from the diffusion-weighted magnetic resonance scans, and extracted the information on electrical potential from the FE ECT model of the same head. We then calculated the first and second spatial derivatives of the electrical potential, which describes the activating function for homogenous axons and investigated sensitive regions of white matter activation.

Recent research reveal that the Working Memory (WM) is more powerful than IQ as a predictor of academic success. However, there are factors that may influence WM performance, such as Attention. Although the impact of attention is well documented using ERPs; yet, the underlying brain connectivity of the interaction of these two constructs is not sufficiently understood. In this study, a Delay-Response task and electroencephalography (EEG) data are used to investigate the brain connectivity during two stages of Working Memory: Encoding and Maintenance. We have presented distraction in both stages, and a secondary task in maintenance stage. Scalp EEG data of 19 participants were recorded. These results not only reveal the underlying brain connectivity of each task, but also highlights the differences between distraction and multitasking. The results show significant brain connectivity changes in the frontal and occipital areas of the brain depending on the WM stage where the distraction is presented.