Recurrent neural networks (RNNs) are often used to model circuits in the brain and can solve a variety of difficult computational problems requiring memory, error correction, or selection (Hopfield, 1982; Maass etal., 2002; Maass, 2011). However, fully connected RNNs contrast structurally with their biological counterparts, which are extremely sparse (about 0.1%). Motivated by the neocortex, where neural connectivity is constrained by physical distance along cortical sheets and other synaptic wiring costs, we introduce locality masked RNNs (LM-RNNs) that use task-agnostic predetermined graphs with sparsity as low as 4%. We study LM-RNNs in a multitask learning setting relevant to cognitive systems neuroscience with a commonly used set of tasks, 20-Cog-tasks (Yang etal., 2019). We show through reductio ad absurdum that 20-Cog-tasks can be solved by a small pool of separated autapses that we can mechanistically analyze and understand. Thus, these tasks fall short of the goal of inducing complex recurrent dynamics and modular structure in RNNs. We next contribute a new cognitive multitask battery, Mod-Cog, consisting of up to 132 tasks that expands by about seven-fold the number of tasks and task complexity of 20-Cog-tasks. Importantly, while autapses can solve the simple 20-Cog-tasks, the expanded task set requires richer neural architectures and continuous attractor dynamics. On these tasks, we show that LM-RNNs with an optimal sparsity result in faster training and better data efficiency than fully connected networks.

Spiking neural networks (SNNs) have emerged as a promising alternative to traditional deep neural networks for low-power computing. However, the effectiveness of SNNs is not solely determined by their performance but also by their energy consumption, prediction speed, and robustness to noise. The recent method Fast & Deep, along with others, achieves fast and energy-efficient computation by constraining neurons to fire at most once. Known as time-to-first-spike (TTFS), this constraint, however, restricts the capabilities of SNNs in many aspects. In this work, we explore the relationships of performance, energy consumption, speed, and stability when using this constraint. More precisely, we highlight the existence of trade-offs where performance and robustness are gained at the cost of sparsity and prediction latency. To improve these trade-offs, we propose a relaxed version of Fast & Deep that allows for multiple spikes per neuron. Our experiments show that relaxing the spike constraint provides higher performance while also benefiting from faster convergence, similar sparsity, comparable prediction latency, and better robustness to noise compared to TTFS SNNs. By highlighting the limitations of TTFS and demonstrating the advantages of unconstrained SNNs, we provide valuable insight for the development of effective learning strategies for neuromorphic computing.

Hopfield networks and Boltzmann machines (BMs) are fundamental energy-based neural network models. Recent studies on modern Hopfield networks have broadened the class of energy functions and led to a unified perspective on general Hopfield networks, including an attention module. In this letter, we consider the BM counterparts of modern Hopfield networks using the associated energy functions and study their salient properties from a trainability perspective. In particular, the energy function corresponding to the attention module naturally introduces a novel BM, which we refer to as the attentional BM (AttnBM). We verify that AttnBM has a tractable likelihood function and gradient for certain special cases and is easy to train. Moreover, we reveal the hidden connections between AttnBM and some single-layer models, namely the gaussian-Bernoulli restricted BM and the denoising autoencoder with softmax units coming from denoising score matching. We also investigate BMs introduced by other energy functions and show that the energy function of dense associative memory models gives BMs belonging to exponential family harmoniums.

Solving the eigenvalue problem for differential operators is a common problem in many scientific fields. Classical numerical methods rely on intricate domain discretization and yield nonanalytic or nonsmooth approximations. We introduce a novel neural network-based solver for the eigenvalue problem of differential self-adjoint operators, where the eigenpairs are learned in an unsupervised end-to-end fashion. We propose several training procedures for solving increasingly challenging tasks toward the general eigenvalue problem. The proposed solver is capable of finding the M smallest eigenpairs for a general differential operator. We demonstrate the method on the Laplacian operator, which is of particular interest in image processing, computer vision, and shape analysis among many other applications. In addition, we solve the Legendre differential equation. Our proposed method simultaneously solves several eigenpairs and can be easily used on free-form domains. We exemplify it on L-shape and circular cut domains. A significant contribution of this work is an analysis of the numerical error of this method. In particular an upper bound for the (unknown) solution error is given in terms of the (measured) truncation error of the partial differential equation and the network structure.

Capsule networks (see Hinton etal., 2018) aim to encode knowledge of and reason about the relationship between an object and its parts. In this letter, we specify a generative model for such data and derive a variational algorithm for inferring the transformation of each model object in a scene and the assignments of observed parts to the objects. We derive a learning algorithm for the object models, based on variational expectation maximization (Jordan etal., 1999). We also study an alternative inference algorithm based on the RANSAC method of Fischler and Bolles (1981). We apply these inference methods to data generated from multiple geometric objects like squares and triangles ("constellations") and data from a parts-based model of faces. Recent work by Kosiorek etal. (2019) has used amortized inference via stacked capsule autoencoders to tackle this problem; our results show that we significantly outperform them where we can make comparisons (on the constellations data).

Deep brain stimulation (DBS) is a widely accepted treatment for the Parkinson's disease (PD). Traditionally, it is done in an open-loop manner, where stimulation is always ON, irrespective of the patient needs. As a consequence, patients can feel some side effects due to the continuous high-frequency stimulation. Closed-loop DBS can address this problem as it allows adjusting stimulation according to the patient need. The selection of open- or closed-loop DBS and an optimal algorithm for closed-loop DBS are some of the main challenges in DBS controller design, and typically the decision is made through sampling based simulations. In this letter, we used model checking, a formal verification technique used to exhaustively explore the complete state space of a system, for analyzing DBS controllers. We analyze the timed automata of the open-loop and closed-loop DBS controllers in response to the basal ganglia (BG) model. Furthermore, we present a formal verification approach for the closed-loop DBS controllers using timed computation tree logic (TCTL) properties, that is, safety, liveness (the property that under certain conditions, some event will eventually occur), and deadlock freeness. We show that the closed-loop DBS significantly outperforms existing open-loop DBS controllers in terms of energy efficiency. Moreover, we formally analyze the closed-loop DBS for energy efficiency and time behavior with two algorithms, the constant update algorithm and the error prediction update algorithm. Our results demonstrate that the closed-loop DBS running the error prediction update algorithm is efficient in terms of time and energy as compared to the constant update algorithm.

How are visuomotor mismatch responses in primary visual cortex embedded into cortical processing? We here show that mismatch responses can be understood as the result of a cooperation of motor and visual areas to jointly explain optic flow. This cooperation requires that optic flow is not explained redundantly by both areas, meaning that optic flow inputs to V1 that are predictable from motor neurons should be canceled (i.e., explained away). As a result, neurons in V1 represent only external causes of optic flow, which could allow the animal to easily detect movements that are independent of its own locomotion. We implement the proposed model in a spiking neural network, where coding errors are computed in dendrites and synaptic weights are learned with voltage-dependent plasticity rules. We find that both positive and negative mismatch responses arise, providing an alternative to the prevailing idea that visuomotor mismatch responses are linked to dedicated neurons for error computation. These results also provide a new perspective on several other recent observations of cross-modal neural interactions in cortex.

Many animal behaviors require orientation and steering with respect to the environment. For insects, a key brain area involved in spatial orientation and navigation is the central complex. Activity in this neural circuit has been shown to track the insect's current heading relative to its environment and has also been proposed to be the substrate of path integration. However, it remains unclear how the output of the central complex is integrated into motor commands. Central complex output neurons project to the lateral accessory lobes (LAL), from which descending neurons project to thoracic motor centers. Here, we present a computational model of a simple neural network that has been described anatomically and physiologically in the LALs of male silkworm moths, in the context of odor-mediated steering. We present and analyze two versions of this network, one rate based and one based on spiking neurons. The modeled network consists of an inhibitory local interneuron and a bistable descending neuron (flip-flop) that both receive input in the LAL. The flip-flop neuron projects onto neck motor neurons to induce steering. We show that this simple computational model not only replicates the basic parameters of male silkworm moth behavior in a simulated odor plume but can also take input from a computational model of path integration in the central complex and use it to steer back to a point of origin. Furthermore, we find that increasing the level of detail within the model improves the realism of the model's behavior, leading to the emergence of looping behavior as an orientation strategy. Our results suggest that descending neurons originating in the LALs, such as flip-flop neurons, are sufficient to mediate multiple steering behaviors. This study is therefore a first step to close the gap between orientation circuits in the central complex and downstream motor centers.

Branching time active inference is a framework proposing to look at planning as a form of Bayesian model expansion. Its root can be found in active inference, a neuroscientific framework widely used for brain modeling, as well as in Monte Carlo tree search, a method broadly applied in the reinforcement learning literature. Up to now, the inference of the latent variables was carried out by taking advantage of the flexibility offered by variational message passing, an iterative process that can be understood as sending messages along the edges of a factor graph. In this letter, we harness the efficiency of an alternative method for inference, Bayesian filtering, which does not require the iteration of the update equations until convergence of the variational free energy. Instead, this scheme alternates between two phases: integration of evidence and prediction of future states. Both phases can be performed efficiently, and this provides a forty times speedup over the state of the art.

Many machine learning methods assume that the training and test data follow the same distribution. However, in the real world, this assumption is often violated. In particular, the marginal distribution of the data changes, called covariate shift, is one of the most important research topics in machine learning. We show that the well-known family of covariate shift adaptation methods is unified in the framework of information geometry. Furthermore, we show that parameter search for a geometrically generalized covariate shift adaptation method can be achieved efficiently. Numerical experiments show that our generalization can achieve better performance than the existing methods it encompasses.