The rapidly growing computational demands of deep neural networks require novel hardware designs. Recently, tunable nanoelectronic devices were developed based on hopping electrons through a network of dopant atoms in silicon. These "Dopant Network Processing Units" (DNPUs) are highly energy-efficient and have potentially very high throughput. By adapting the control voltages applied to its terminals, a single DNPU can solve a variety of linearly non-separable classification problems. However, using a single device has limitations due to the implicit single-node architecture. This paper presents a promising novel approach to neural information processing by introducing DNPUs as high-capacity neurons and moving from a single to a multi-neuron framework. By implementing and testing a small multi-DNPU classifier in hardware, we show that feed-forward DNPU networks improve the performance of a single DNPU from 77% to 94% test accuracy on a binary classification task with concentric classes on a plane. Furthermore, motivated by the integration of DNPUs with memristor arrays, we study the potential of using DNPUs in combination with linear layers. We show by simulation that a single-layer MNIST classifier with only 10 DNPUs achieves over 96% test accuracy. Our results pave the road towards hardware neural-network emulators that offer atomic-scale information processing with low latency and energy consumption.Preprint. Under review.

Non-invasive methods to measure brain activity are important to understand cognitive processes in the human brain. A prominent example is functional magnetic resonance imaging (fMRI), which is a noisy measurement of a delayed signal that depends non-linearly on the neuronal activity through the neurovascular coupling. These characteristics make the inference of neuronal activity from fMRI a difficult but important step in fMRI studies that require information at the neuronal level. In this article, we address this inference problem using a Bayesian approach where we model the latent neural activity as a stochastic process and assume that the observed BOLD signal results from a realistic physiological (Balloon) model. We apply a recently developed smoothing method called APIS to efficiently sample the posterior given single event fMRI time series. To infer neuronal signals with high likelihood for multiple time series efficiently, a modification of the original algorithm is introduced. We demonstrate that our adaptive procedure is able to compensate the lacking of inputs in the model to infer the neuronal activity and that it outperforms dramatically the standard bootstrap particle filter-smoother in this setting. This makes the proposed procedure specially attractive to deconvolve resting state fMRI data. To validate the method, we evaluate the quality of the signals inferred using the timing information contained in them. APIS obtains reliable event timing estimates based on fMRI data gathered during a reaction time experiment with short stimuli. Hence, we show for the first time that one can obtain accurate absolute timing of neuronal activity by reconstructing the latent neural signal.