Mijung Park

Receptive Field Inference with Localized Priors

The linear receptive field describes a mapping from sensory stimuli to a one-dimensional variable governing a neurons spike response. However, traditional receptive field estimators such as the spike-triggered average converge slowly and often require large amounts of data. Bayesian methods seek to overcome this problem by biasing estimates towards solutions that are more likely a priori, typically those with small, smooth, or sparse coefficients. Here we introduce a novel Bayesian receptive field estimator designed to incorporate locality, a powerful form of prior information about receptive field structure. The key to our approach is a hierarchical receptive field model that flexibly adapts to localized structure in both spacetime and spatiotemporal frequency, using an inference method known as empirical Bayes. We refer to our method as automatic locality determination (ALD), and show that it can accurately recover various types of smooth, sparse, and localized receptive fields. We apply ALD to neural data from retinal ganglion cells and V1 simple cells, and find it achieves error rates several times lower than standard estimators. Thus, estimates of comparable accuracy can be achieved with substantially less data. Finally, we introduce a computationally efficient Markov Chain Monte Carlo (MCMC) algorithm for fully Bayesian inference under the ALD prior, yielding accurate Bayesian confidence intervals for small or noisy datasets.

Bayesian Active Learning for Drug Combinations

The dynamics of complex diseases are governed by intricate interactions of myriad factors. Drug combinations, formed by mixing several single-drug treatments at various doses, can enhance the effectiveness of the therapy by targeting multiple contributing factors. The main challenge in designing drug combinations is the highly nonlinear interaction of the constituent drugs. Prior work focused on guided space-exploratory heuristics that require discretization of drug doses. While being more efficient than random sampling, these methods are impractical if the drug space is high dimensional or if the drug sensitivity is unknown. Furthermore, the effectiveness of the obtained combinations may decrease if the resolution of the discretization grid is not sufficiently fine. In this paper, we model the biological system response to a continuous combination of drug doses by a Gaussian process (GP). We perform closed-loop experiments that rely on the expected improvement criterion to efficiently guide the exploration process toward drug combinations with the optimal response. When computing the criterion, we marginalize out the GP hyperparameters in a fully Bayesian manner using a particle filter. Finally, we employ a hybrid Monte Carlo algorithm to rapidly explore the high-dimensional continuous search space. We demonstrate the effectiveness of our approach on a fully factorial Drosophila dataset, an antiviral drug dataset for Herpes simplex virus type 1, and simulated human Apoptosis networks. The results show that our approach significantly reduces the number of required trials compared to existing methods.

Adaptive Bayesian Methods for Closed-Loop Neurophysiology

An important problem in the design of neurophysiology experiments is to select stimuli that rapidly probe a neuron’s tuning or response properties. This is especially important in settings where the neural parameter space is multidimensional and the experiment is limited in time. Bayesian active learning methods provide a formal solution to this problem using a statistical model of the neural response and a utility function that quantifies what we want to learn. In contrast to staircase and other ad hoc stimulus selection methods, Bayesian active learning methods use the entire set of past stimuli and responses to make inferences about functional properties and select the next stimulus. Here we discuss recent advances in Bayesian active learning methods for closed-loop neurophysiology experiments. We review the general ingredients for Bayesian active learning and then discuss two specific applications in detail: (1) low-dimensional nonlinear response surfaces (also known as “tuning curves” or “firing rate maps”) and (2) high-dimensional linear receptive fields. Recent work has shown that these methods can achieve higher accuracy in less time, allowing for experiments that are infeasible with nonadaptive methods. We conclude with a discussion of open problems and exciting directions for future research.

Bayesian active learning of neural firing rate maps with transformed gaussian process priors

A firing rate map, also known as a tuning curve, describes the nonlinear relationship between a neurons spike rate and a low-dimensional stimulus (e.g., orientation, head direction, contrast, color). Here we investigate Bayesian active learning methods for estimating firing rate maps in closed-loop neurophysiology experiments. These methods can accelerate the characterization of such maps through the intelligent, adaptive selection of stimuli. Specifically, we explore the manner in which the prior and utility function used in Bayesian active learning affect stimulus selection and performance. Our approach relies on a flexible model that involves a nonlinearly transformed gaussian process (GP) prior over maps and conditionally Poisson spiking. We show that infomax learning, which selects stimuli to maximize the information gain about the firing rate map, exhibits strong dependence on the seemingly innocuous choice of nonlinear transformation function. We derive an alternate utility function that selects stimuli to minimize the average posterior variance of the firing rate map and analyze the surprising relationship between prior parameterization, stimulus selection, and active learning performance in GP-Poisson models. We apply these methods to color tuning measurements of neurons in macaque primary visual cortex.

A Machine Learning Approach to Link Adaptation for SC-FDE System

Single carrier frequency domain equalization (SC-FDE) uses cyclically prefixed quadrature amplitude modulation to permit simple frequency domain equalization at the receiver. Link adaptation for SC-FDE systems, where the modulation and coding rate are adapted based on the current channel state, is straightforward with perfect channel state information due to the simple analytical form of the post-processing signal-to-noise ratio (SNR). Imperfect channel state information, however, introduces adaptation errors. This paper proposes a machine learning-based approach for link adaptation in bit interleaved convolutionally encoded SC-FDE systems. To improve performance in the presence of channel uncertainty, principal component analysis is used to reduce the feature space dimensionality consisting of the channel coefficients, noise variance, and post-processing SNR. The reduced dimension feature set improves performance of the link adaptation classifier and leads to higher performance versus just the post-processing SNR estimate. Simulation results indicate that the proposed algorithm increases the goodput while maintaining the target packet error rate, achieving optimal adaptation in 95% of the tested cases.

Adaptive experimental design for drug combinations

Drug cocktails formed by mixing multiple drugs at various doses provide more effective cures than single-drug treatments. However, drugs interact in highly nonlinear ways making the determination of the optimal combination a difficult task. The response surface of the drug cocktail has to be estimated through expensive and time-consuming experimentation. Previous research focused on the use of space-exploratory heuristics such as genetic algorithms to guide the search for optimal combinations. While being more efficient than random sampling, these methods require a considerable amount of experiments to converge to good solutions. In this paper, we propose to use an information-theoretic active learning approach under the Bayesian framework of Gaussian processes to adaptively choose what experiments to perform based on current data points. We show that our approach is able to reduce the number of required data points significantly.