Person:

Maclaurin, Dougal

We developed a technique, “flash memory”, to record a photochemical imprint of the activity state—firing or not firing—of a neuron at a user-selected moment in time. The key element is an engineered microbial rhodopsin protein with three states. Two nonfluorescent states, D1 and D2, exist in a voltage-dependent equilibrium. A stable fluorescent state, F, is reached by a photochemical conversion from D2. When exposed to light of a wavelength λwrite, population transfers from D2 to F, at a rate determined by the D1 ⇌ D2 equilibrium. The population of F maintains a record of membrane voltage which persists in the dark. Illumination at a later time at a wavelength λread excites fluorescence of F, probing this record. An optional third flash at a wavelength λreset converts F back to D2, for a subsequent write–read cycle. The flash memory method offers the promise to decouple the recording of neural activity from its readout. In principle, the technique may enable one to generate snapshots of neural activity in a large volume of neural tissue, e.g., a complete mouse brain, by circumventing the challenge of imaging a large volume with simultaneous high spatial and high temporal resolution. The proof-of-principle flash memory sensors presented here will need improvements in sensitivity, speed, brightness, and membrane trafficking before this goal can be realized.

Frequency standards based on atomic states, such as Rb or Cs vapors, or single-trapped ions, are the most precise measures of time. Here we propose and analyze a precision oscillator approach based upon spins in a solid-state system, in particular, the nitrogen-vacancy defect in single-crystal diamond. We show that this system can have stability approaching portable atomic standards and is readily incorporable as a chip-scale device. Using a pulsed spin-echo technique, we anticipate an Allan deviation of σy=10−7τ−1/2 limited by thermally-induced strain variations; in the absence of such thermal fluctuations, the system is limited by spin dephasing and harbors an Allan deviation nearing ∼10−12τ−1/2. Potential improvements based upon advanced diamond material processing, temperature stabilization, and nanophotonic engineering are discussed.

We introduce a convolutional neural network that operates directly on graphs. These networks allow end-to-end learning of prediction pipelines whose inputs are graphs of arbitrary size and shape. The architecture we present generalizes standard molecular feature extraction methods based on circular fingerprints. We show that these data-driven features are more interpretable, and have better predictive performance on a variety of tasks.

All-optical electrophysiology—spatially resolved simultaneous optical perturbation and measurement of membrane voltage—would open new vistas in neuroscience research. We evolved two archaerhodopsin-based voltage indicators, QuasAr1 and 2, which show improved brightness and voltage sensitivity, microsecond response times, and produce no photocurrent. We engineered a novel channelrhodopsin actuator, CheRiff, which shows improved light sensitivity and kinetics, and spectral orthogonality to the QuasArs. A co-expression vector, Optopatch, enabled crosstalk-free genetically targeted all-optical electrophysiology. In cultured neurons, we combined Optopatch with patterned optical excitation to probe back-propagating action potentials in dendritic spines, synaptic transmission, sub-cellular microsecond-timescale details of action potential propagation, and simultaneous firing of many neurons in a network. Optopatch measurements revealed homeostatic tuning of intrinsic excitability in human stem cell-derived neurons. In brain slice, Optopatch induced and reported action potentials and subthreshold events, with high signal-to-noise ratios. The Optopatch platform enables high-throughput, spatially resolved electrophysiology without use of conventional electrodes.

This thesis presents five contributions to machine learning, with themes of differentiability and Bayesian inference.

We present Firefly Monte Carlo, an auxiliary variable Markov chain Monte Carlo algorithm that only queries a potentially small subset of data at each iteration yet simulates from the exact posterior distribution.

We describe the design and implementation of Autograd, a software package for efficiently computing derivatives of functions written in Python/Numpy using reverse accumulation mode differentiation.

Using Autograd, we develop a convolutional neural network that takes arbitrary graphs, such as organic molecules, as input. This generalizes standard molecular feature representations and allows end-to-end adaptation of the feature extraction pipeline to particular tasks.

We show how to compute gradients of cross-validation loss with respect to hyperparameters of learning algorithms, with both time and memory efficiency, by chaining gradients backwards through an exactly reversed optimization procedure.

Finally, by accounting for the entropy destroyed by optimization, we show that early stopping and ensembling, popular tricks for avoiding overfitting, can be interpreted as variational Bayesian inference.

Markov chain Monte Carlo (MCMC) is a popular and successful general-purpose tool for Bayesian inference. However, MCMC cannot be practically applied to large data sets because of the prohibitive cost of evaluating every likelihood term at every iteration. Here we present Firefly Monte Carlo (FlyMC) an auxiliary variable MCMC algorithm that only queries the likelihoods of a potentially small subset of the data at each iteration yet simulates from the exact posterior distribution, in contrast to recent proposals that are approximate even in the asymptotic limit. FlyMC is compatible with a wide variety of modern MCMC algorithms, and only requires a lower bound on the per-datum likelihood factors. In experiments, we find that FlyMC generates samples from the posterior more than an order of magnitude faster than regular MCMC, opening up MCMC methods to larger datasets than were previously considered feasible.