Targeted free energy estimation via learned mappings.

Free energy perturbation (FEP) was proposed by Zwanzig [J. Chem. Phys. 22, 1420 (1954)] more than six decades ago as a method to estimate free energy differences and has since inspired a huge body of related methods that use it as an integral building block. Being an importance sampling based estimator, however, FEP suffers from a severe limitation: the requirement of sufficient overlap between distributions. One strategy to mitigate this problem, called Targeted FEP, uses a high-dimensional mapping in configuration space to increase the overlap of the underlying distributions. Despite its potential, this method has attracted only limited attention due to the formidable challenge of formulating a tractable mapping. Here, we cast Targeted FEP as a machine learning problem in which the mapping is parameterized as a neural network that is optimized so as to increase the overlap. We develop a new model architecture that respects permutational and periodic symmetries often encountered in atomistic simulations and test our method on a fully periodic solvation system. We demonstrate that our method leads to a substantial variance reduction in free energy estimates when compared against baselines, without requiring any additional data.

Inference of perceptual priors from path dynamics of passive self-motion.

The monitoring of one's own spatial orientation depends on the ability to estimate successive self-motion cues accurately. This process has become to be known as path integration. A feature of sequential cue estimation, in general, is that the history of previously experienced stimuli, or priors, biases perception. Here, we investigate how during angular path integration, the prior imparted by the displacement path dynamics affects the translation of vestibular sensations into perceptual estimates. Subjects received successive whole-body yaw rotations and were instructed to report their position within a virtual scene after each rotation. The overall movement trajectory either followed a parabolic path or was devoid of explicit dynamics. In the latter case, estimates were biased toward the average stimulus prior and were well captured by an optimal Bayesian estimator model fit to the data. However, the use of parabolic paths reduced perceptual uncertainty, and a decrease of the average size of bias and thus the weight of the average stimulus prior were observed over time. The produced estimates were, in fact, better accounted for by a model where a prediction of rotation magnitude is inferred from the underlying path dynamics on each trial. Therefore, when passively displaced, we seem to be able to build, over time, from sequential vestibular measurements an internal model of the vehicle's movement dynamics. Our findings suggest that in ecological conditions, vestibular afference can be internally predicted, even when self-motion is not actively generated by the observer, thereby augmenting both the accuracy and precision of displacement perception.

Neural scene representation and rendering.

Scene representation-the process of converting visual sensory data into concise descriptions-is a requirement for intelligent behavior. Recent work has shown that neural networks excel at this task when provided with large, labeled datasets. However, removing the reliance on human labeling remains an important open problem. To this end, we introduce the Generative Query Network (GQN), a framework within which machines learn to represent scenes using only their own sensors. The GQN takes as input images of a scene taken from different viewpoints, constructs an internal representation, and uses this representation to predict the appearance of that scene from previously unobserved viewpoints. The GQN demonstrates representation learning without human labels or domain knowledge, paving the way toward machines that autonomously learn to understand the world around them.

Stochastic variational learning in recurrent spiking networks.

The ability to learn and perform statistical inference with biologically plausible recurrent networks of spiking neurons is an important step toward understanding perception and reasoning. Here we derive and investigate a new learning rule for recurrent spiking networks with hidden neurons, combining principles from variational learning and reinforcement learning. Our network defines a generative model over spike train histories and the derived learning rule has the form of a local Spike Timing Dependent Plasticity rule modulated by global factors (neuromodulators) conveying information about "novelty" on a statistically rigorous ground. Simulations show that our model is able to learn both stationary and non-stationary patterns of spike trains. We also propose one experiment that could potentially be performed with animals in order to test the dynamics of the predicted novelty signal.