The human brain and modern computers are both extremely efficient computational machines, executing large amounts of operations at incredibly fast speeds. Such performance costs energy, with the theoretical minimum being Landauer’s limit. We attempt to investigate and compare the efficiency of computation between our brain and the modern computer, i.e. cost of energy per bit of information. Every machine, including our brain, specialize in certain types of computation. We are more adept at survival tasks such as spotting colors versus mathematical operators, which the modern computer is so efficient at. Therefore, we will avoid the algorithm and architecture aspect of these computational machines. We instead quantify what a bit of information is and estimate the energy cost in both machines. We find that modern computers are extremely efficient, requiring approximately 10-19 J/bit which is only 103 away from the Landaeur’s limit while our brain requires 10-18 J/bit. Despite the lack of precision, the disparity is great enough to establish that modern computers are more efficient at executing a bit of computation.

Quantum Machine Learning is a rapidly growing field in the applied and theoretical machine learning community. A number of novel techniques are proposed for integrating quantum machine learning into classical machine learning methods and paradigms, including quantum self-attention in Vision Transformers, using parameterized unimodular matrices for quantum-classical hybrid machine learning models, and various ways of inserting quantum machine learning into spiking neural networks. The above techniques are trained and tested on standardized datasets used within the machine learning community. All techniques presented are evaluated​ against comparable classical models​.

There has been a recent trend to applying deep learning methods compared to shallowmethods for automatic identification of insects. Classification strategies built around al- gorithms with deep learning architectures at their center like YOLO and others require large amounts of data to making learning successful and are often augmented with tens of thousands of images or more to achieve excellent performance. Recent pre-trained models of deep neural networks have significantly reduced the amount of data required to create accurate classification algorithms by ingesting and training on a huge data set different than the target task and using the resulting encoding to transfer information to a new task. This work shows that recent performance gains from models pre-trained on huge data sets are effective as image encoders for the classification of the sex of spotted wing drosophila (SWD). A data set of 676 SWD microscope images is created to evaluate classification models for use in automation of the sterile insect technique (SIT), which requires large amounts of male SWD to be identified and separated. Bi- nary classification models trained on top of image encoding from new models based off of visual transformers [3] pre-trained on over 400 million images with CLIP [2] are able to achieve accuracy as high as 96.7% when trained with LogReg and similar classifiers on augmented data from the SWD image set. Other models pre-trained on the ImageNet data set of 14 million images also performed well, approaching 92% with VGG models and 90% with MobileNetV2 model. Image segmentation of the data set is then inves- tigated as a source of corroboration for the identification of the morphological features responsible for classification, and an out-of-distribution data set is collected to evaluate classification and segmentation results on more diverse and difficult examples. While robust identification of features special to SWD remains, classification accuracy is not a guarantee on data which differs substantially from the factory or laboratory setting on which it is trained and additional data may be needed for training on use-cases outside of SIT such as for applications on the farm or for automated identification in insect traps. This emphasizes a fact which is not elaborated on for many insect detection models in the literature: that their models are not likely robust in situations where the data is significantly OOD and for situations which may not be adequately covered with- out specialized augmentation methods or additional data. Nonetheless results indicate that pre-trained models have advanced to the point where they can play a central role in securing the food supply from potentially billions of dollars of damages every year from pests such as SWD.

Neural networks are a powerful tool which have shown usefulness in a number of challenging domains. We have developed novel methods which incorporate neural networks in two important domains. The first domain considered is in mortality prediction in intensive care units. For this problem, a model involving long short term memory recurrent neural networks was developed. This model was trained and tested using the Medical Information Mart for Intensive Care (MIMIC iii) data set. Our method generates a mortality trajectory after each input health measurement consisting of a sequence of mortality predictions. Through testing we demonstrate that our model is capable of providing effective mortality predictions in this data set. The next domain considered is production scheduling in semiconductor manufacturing facilities. We developed two methods for training scheduling policies that utilize deep reinforcement learning. The first involves a Deep Q-Network, and the second is a novel method we developed known as the Predictron Deep Q-Network. A factory simulation model was developed. Our methods were demonstrated to outperform common industry methods on a set of simulated factory environments based on real world systems.

Free-Electron Lasers like the one at the SLAC National Accelerator Laboratory are sources of extremely bright X-rays that are useful in a variety of scientific imaging applications. Because there are only a handful of FEL facilities around the world, access to these X-rays is in high demand. Every year, hundreds of hours are spent tuning quadrupole focusing magnets to optimize the X-ray brightness. During this tuning process, the beam typically cannot be used for experiments. In this thesis, I show that by performing Bayesian optimization using a Gaussian process regression model containing prior information derived from an optical model of the accelerator in combination with historical data, we were able to significantly reduce the amount of time spent tuning the quadrupoles in comparison to previous methods.

Three active matter systems involving long, flexible molecules (which we often refer to generally as polymers) are examined. Phenomena arising from these systems -- be it pure prediction (as in chapters 2 and 4) or based on a known experiment (chapter 3) -- are analyzed theoretically and through simulation. The first system we consider is a plate resting on a tilted polymer brush, where the ends of the polymers bind to sites on the plate and unbind in response to an incoming photon. We show that, under the right conditions, the plate will move in the direction of the polymer tilt, converting incoming light energy directly to mechanical energy. This mechanism was inspired by a similar mechanism already present in biology: the actin-myosin interface in muscle tissue. The second system is that of microtubule bundles confined between two parallel plates with one end bound to a surface, as studied experimentally by Sanchez et al. In this experiment, the spontaneous formation of metachronal waves was observed. We show that this result can be reproduced using first-principles hydrodynamic simulations, and argue that these basic interactions are sufficient to explain the observed behavior. Finally, we propose an experiment in which microtubules are not bound to a surface, but rather confined to an optical trap. We show that, under the right conditions, drifting microtubules exhibit various phases, and can even align, circling the trap uniformly.