arXiv subjects

Kyle Mills

Publications and source records attributed to Kyle Mills.

9 recordsLinked to original sources

arXiv2021

Weakly-supervised multi-class object localization using only object counts as labels

We demonstrate the use of an extensive deep neural network to localize instances of objects in images. The EDNN is naturally able to accurately perform multi-class counting using only ground truth count values as labels. Without providing any conceptual information, object annotations, or pixel segmentation information, the neural network is able to formulate its own conceptual representation of the items in the image. Using images labelled with only the counts of the objects present,the structure of the extensive deep neural network can be exploited to perform localization of the objects within the visual field. We demonstrate that a trained EDNN can be used to count objects in images much larger than those on which it was trained. In order to demonstrate our technique, we introduce seven new data sets: five progressively harder MNIST digit-counting data sets, and two datasets of 3d-rendered rubber ducks in various situations. On most of these datasets, the EDNN achieves greater than 99% test set accuracy in counting objects.

cs.CV↗

arXiv2020

Controlled Online Optimization Learning (COOL): Finding the ground state of spin Hamiltonians with reinforcement learning

Reinforcement learning (RL) has become a proven method for optimizing a procedure for which success has been defined, but the specific actions needed to achieve it have not. We apply the so-called "black box" method of RL to what has been referred as the "black art" of simulated annealing (SA), demonstrating that an RL agent based on proximal policy optimization can, through experience alone, arrive at a temperature schedule that surpasses the performance of standard heuristic temperature schedules for two classes of Hamiltonians. When the system is initialized at a cool temperature, the RL agent learns to heat the system to "melt" it, and then slowly cool it in an effort to anneal to the ground state; if the system is initialized at a high temperature, the algorithm immediately cools the system. We investigate the performance of our RL-driven SA agent in generalizing to all Hamiltonians of a specific class; when trained on random Hamiltonians of nearest-neighbour spin glasses, the RL agent is able to control the SA process for other Hamiltonians, reaching the ground state with a higher probability than a simple linear annealing schedule. Furthermore, the scaling performance (with respect to system size) of the RL approach is far more favourable, achieving a performance improvement of one order of magnitude on L=14x14 systems. We demonstrate the robustness of the RL approach when the system operates in a "destructive observation" mode, an allusion to a quantum system where measurements destroy the state of the system. The success of the RL agent could have far-reaching impact, from classical optimization, to quantum annealing, to the simulation of physical systems.

physics.comp-ph↗

arXiv2020

Optical lattice experiments at unobserved conditions and scales through generative adversarial deep learning

Machine learning provides a novel avenue for the study of experimental realizations of many-body systems, and has recently been proven successful in analyzing properties of experimental data of ultracold quantum gases. We here show that deep learning succeeds in the more challenging task of modelling such an experimental data distribution. Our generative model (RUGAN) is able to produce snapshots of a doped two-dimensional Fermi-Hubbard model that are indistinguishable from previously reported experimental realizations. Importantly, it is capable of accurately generating snapshots at conditions for which it did not observe any experimental data, such as at higher doping values. On top of that, our generative model extracts relevant patterns from small-scale examples and can use these to construct new configurations at a larger size that serve as a precursor to observations at scales that are currently experimentally inaccessible. The snapshots created by our model---which come at effectively no cost---are extremely useful as they can be employed to quantitatively test new theoretical developments under conditions that have not been explored experimentally, parameterize phenomenological models, or train other, more data-intensive, machine learning methods. We provide predictions for experimental observables at unobserved conditions and benchmark these against modern theoretical frameworks. The deep learning method we develop here is broadly applicable and can be used for the efficient large-scale simulation of equilibrium and nonequilibrium physical systems.

physics.comp-ph↗

arXiv2019

Optimizing thermodynamic trajectories using evolutionary and gradient-based reinforcement learning

Using a model heat engine, we show that neural network-based reinforcement learning can identify thermodynamic trajectories of maximal efficiency. We consider both gradient and gradient-free reinforcement learning. We use an evolutionary learning algorithm to evolve a population of neural networks, subject to a directive to maximize the efficiency of a trajectory composed of a set of elementary thermodynamic processes; the resulting networks learn to carry out the maximally-efficient Carnot, Stirling, or Otto cycles. When given an additional irreversible process, this evolutionary scheme learns a previously unknown thermodynamic cycle. Gradient-based reinforcement learning is able to learn the Stirling cycle, whereas an evolutionary approach achieves the optimal Carnot cycle. Our results show how the reinforcement learning strategies developed for game playing can be applied to solve physical problems conditioned upon path-extensive order parameters.

cs.NE↗

arXiv2017

Deep learning and the Schrödinger equation

We have trained a deep (convolutional) neural network to predict the ground-state energy of an electron in four classes of confining two-dimensional electrostatic potentials. On randomly generated potentials, for which there is no analytic form for either the potential or the ground-state energy, the neural network model was able to predict the ground-state energy to within chemical accuracy, with a median absolute error of 1.49 mHa. We also investigate the performance of the model in predicting other quantities such as the kinetic energy and the first excited-state energy of random potentials.

cond-mat.mtrl-sci↗

arXiv2017

Phase space sampling and operator confidence with generative adversarial networks

We demonstrate that a generative adversarial network can be trained to produce Ising model configurations in distinct regions of phase space. In training a generative adversarial network, the discriminator neural network becomes very good a discerning examples from the training set and examples from the testing set. We demonstrate that this ability can be used as an anomaly detector, producing estimations of operator values along with a confidence in the prediction.

cond-mat.stat-mech↗

arXiv2017

Extensive deep neural networks for transferring small scale learning to large scale systems

We present a physically-motivated topology of a deep neural network that can efficiently infer extensive parameters (such as energy, entropy, or number of particles) of arbitrarily large systems, doing so with O(N) scaling. We use a form of domain decomposition for training and inference, where each sub-domain (tile) is comprised of a non-overlapping focus region surrounded by an overlapping context region. The size of these regions is motivated by the physical interaction length scales of the problem. We demonstrate the application of EDNNs to three physical systems: the Ising model and two hexagonal/graphene-like datasets. In the latter, an EDNN was able to make total energy predictions of a 60 atoms system, with comparable accuracy to density functional theory (DFT), in 57 milliseconds. Additionally EDNNs are well suited for massively parallel evaluation, as no communication is necessary during neural network evaluation. We demonstrate that EDNNs can be used to make an energy prediction of a two-dimensional 35.2 million atom system, over 1 square micrometer of material, at an accuracy comparable to DFT, in under 25 minutes. Such a system exists on a length scale visible with optical microscopy and larger than some living organisms.

physics.comp-ph↗

arXiv2017

Deep neural networks for direct, featureless learning through observation: the case of 2d spin models

We demonstrate the capability of a convolutional deep neural network in predicting the nearest-neighbor energy of the 4x4 Ising model. Using its success at this task, we motivate the study of the larger 8x8 Ising model, showing that the deep neural network can learn the nearest-neighbor Ising Hamiltonian after only seeing a vanishingly small fraction of configuration space. Additionally, we show that the neural network has learned both the energy and magnetization operators with sufficient accuracy to replicate the low-temperature Ising phase transition. We then demonstrate the ability of the neural network to learn other spin models, teaching the convolutional deep neural network to accurately predict the long-range interaction of a screened Coulomb Hamiltonian, a sinusoidally attenuated screened Coulomb Hamiltonian, and a modified Potts model Hamiltonian. In the case of the long-range interaction, we demonstrate the ability of the neural network to recover the phase transition with equivalent accuracy to the numerically exact method. Furthermore, in the case of the long-range interaction, the benefits of the neural network become apparent; it is able to make predictions with a high degree of accuracy, and do so 1600 times faster than a CUDA-optimized exact calculation. Additionally, we demonstrate how the neural network succeeds at these tasks by looking at the weights learned in a simplified demonstration.

cond-mat.mtrl-sci↗

arXiv2017

Convolutional neural networks for atomistic systems

We introduce a new method, called CNNAS (convolutional neural networks for atomistic systems), for calculating the total energy of atomic systems which rivals the computational cost of empirical potentials while maintaining the accuracy of \emph{ab initio} calculations. This method uses deep convolutional neural networks (CNNs), where the input to these networks are simple representations of the atomic structure. We use this approach to predict energies obtained using density functional theory (DFT) for 2D hexagonal lattices of various types. Using a dataset consisting of graphene, hexagonal boron nitride (hBN), and graphene-hBN heterostructures, with and without defects, we trained a deep CNN that is capable of predicting DFT energies to an extremely high accuracy, with a mean absolute error (MAE) of 0.198 meV / atom (maximum absolute error of 16.1 meV / atom). To explore our new methodology, we investigate the ability of a deep neural network (DNN) in predicting a Lennard-Jones energy and separation distance for a dataset of dimer molecules in both two and three dimensions. In addition, we systematically investigate the flexibility of the deep learning models by performing interpolation and extrapolation tests.

cond-mat.mtrl-sci↗