Kenneth Chiu

Idempotent generative networks (IGNs) are a new line of generative models based on idempotent mapping to a target manifold. IGNs support both single-and multi-step generation, allowing for a flexible trade-off between computational cost and sample quality. But similar to Generative Adversarial Networks (GANs), conventional IGNs require adversarial training and are prone to training instabilities and mode collapse. Diffusion and score-based models are popular approaches to generative modeling that iteratively transport samples from one distribution, usually a Gaussian, to a target data distribution. These models have gained popularity due to their stable training dynamics and high-fidelity generation quality. However, this stability and quality come at the cost of high computational cost, as the data must be transported incrementally along the entire trajectory. New sampling methods, model distillation, and consistency models have been developed to reduce the sampling cost and even perform one-shot sampling from diffusion models. In this work, we unite diffusion and IGNs by distilling idempotent models from diffusion model scores, called SIGN. Our proposed method is highly stable and does not require adversarial losses. We provide a theoretical analysis of our proposed score-based training methods and empirically show that IGNs can be effectively distilled from a pre-trained diffusion model, enabling faster inference than iterative score-based models. SIGNs can perform multi-step sampling, allowing users to trade off quality for efficiency. These models operate directly on the source domain; they can project corrupted or alternate distributions back onto the target manifold, enabling zero-shot editing of inputs. We validate our models on multiple image datasets, achieving state-of-the-art results for idempotent models on the CIFAR and CelebA datasets.

Ptychography is a computational method of microscopy that recovers high-resolution transmission images of samples from a series of diffraction patterns. While conventional phase retrieval algorithms can iteratively recover the images, they require oversampled diffraction patterns, incur significant computational costs, and struggle to recover the absolute phase of the sample's transmission function. Deep learning algorithms for ptychography are a promising approach to resolving the limitations of iterative algorithms. We present PtychoFormer, a hierarchical transformer-based model for data-driven single-shot ptychographic phase retrieval. PtychoFormer processes subsets of diffraction patterns, generating local inferences that are seamlessly stitched together to produce a high-quality reconstruction. Our model exhibits tolerance to sparsely scanned diffraction patterns and achieves up to 3600 times faster imaging speed than the extended ptychographic iterative engine (ePIE). We also propose the extended-PtychoFormer (ePF), a hybrid approach that combines the benefits of PtychoFormer with the ePIE. ePF minimizes global phase shifts and significantly enhances reconstruction quality, achieving state-of-the-art phase retrieval in ptychography.

Deep reinforcement learning has successfully been applied for molecular discovery as shown by the Molecule Deep Q-network (MolDQN) algorithm. This algorithm has challenges when applied to optimizing new molecules: training such a model is limited in terms of scalability to larger datasets and the trained model cannot be generalized to different molecules in the same dataset. In this paper, a distributed reinforcement learning algorithm for antioxidants, called DA-MolDQN is proposed to address these problems. State-of-the-art bond dissociation energy (BDE) and ionization potential (IP) predictors are integrated into DA-MolDQN, which are critical chemical properties while optimizing antioxidants. Training time is reduced by algorithmic improvements for molecular modifications. The algorithm is distributed, scalable for up to 512 molecules, and generalizes the model to a diverse set of molecules. The proposed models are trained with a proprietary antioxidant dataset. The results have been reproduced with both proprietary and public datasets. The proposed molecules have been validated with DFT simulations and a subset of them confirmed in public "unseen" datasets. In summary, DA-MolDQN is up to 100x faster than previous algorithms and can discover new optimized molecules from proprietary and public antioxidants.

Data processing systems are a fundamental component of the modern computing stack. These systems are routinely deployed online: they continuously receive the requests of data processing operations, and continuously return the results to end users or client applications. Online data processing systems have unique features beyond conventional data processing, and the optimizations designed for them are complex, especially when data themselves are structured and dynamic. This paper describes DON Calculus, the first rigorous foundation for online data processing. It captures the essential behavior of both the backend data processing engine and the frontend application, with the focus on two design dimensions essential yet unique to online data processing systems: incremental operation processing (IOP) and temporal locality optimization (TLO). A novel design insight is that the operations continuously applied to the data can be defined as an operation stream flowing through the data structure, and this abstraction unifies diverse designs of IOP and TLO in one calculus. DON Calculus is endowed with a mechanized metatheory centering around a key observable equivalence property: despite the significant non-deterministic executions introduced by IOP and TLO, the observable result of DON Calculus data processing is identical to that of conventional data processing without IOP and TLO. Broadly, DON Calculus is a novel instance in the active pursuit of providing rigorous guarantees to the software system stack. The specification and mechanization of DON Calculus provide a sound base for the designers of future data processing systems to build upon, helping them embrace rigorous semantic engineering without the need of developing from scratch.

Metal-organic frameworks (MOFs) are nanoporous compounds composed of metal ions and organic linkers. MOFs play an important role in industrial applications such as gas separation, gas purification, and electrolytic catalysis. Important MOF properties such as potential energy are currently computed via techniques such as density functional theory (DFT). Although DFT provides accurate results, it is computationally costly. We propose a machine learning approach for estimating the potential energy of candidate MOFs, decomposing it into separate pair-wise atomic interactions using a graph neural network. Such a technique will allow high-throughput screening of candidates MOFs. We also generate a database of 50,000 spatial configurations and high-quality potential energy values using DFT.

Recent advances in unmanned-aerial-vehicle- (UAV-) based remote sensing utilizing lightweight multispectral and thermal infrared sensors allow for rapid wide-area landmine contamination detection and mapping surveys. We present results of a study focused on developing and testing an automated technique of remote landmine detection and identification of scatterable antipersonnel landmines in wide-area surveys. Our methodology is calibrated for the detection of scatterable plastic landmines which utilize a liquid explosive encapsulated in a polyethylene or plastic body in their design. We base our findings on analysis of multispectral and thermal datasets collected by an automated UAV-survey system featuring scattered PFM-1-type landmines as test objects and present results of an effort to automate landmine detection, relying on supervised learning algorithms using a Faster Regional-Convolutional Neural Network (Faster R-CNN). The RGB visible light Faster R-CNN demo yielded a 99.3% testing accuracy for a partially withheld testing set and 71.5% testing accuracy for a completely withheld testing set. Across multiple test environments, using centimeter scale accurate georeferenced datasets paired with Faster R-CNN, allowed for accurate automated detection of test PFM-1 landmines. This method can be calibrated to other types of scatterable antipersonnel mines in future trials to aid humanitarian demining initiatives. With millions of remnant PFM-1 and similar scatterable plastic mines across post-conflict regions and considerable stockpiles of these landmines posing long-term humanitarian and economic threats to impacted communities, our methodology could considerably aid in efforts to demine impacted regions.

Rigorous study of mitochondrial functions and cell biology in the budding yeast, Saccharomyces cerevisiae has advanced our understanding of mitochondrial genetics. This yeast is now a powerful model for population genetics, owing to large genetic diversity and highly structured populations among wild isolates. Comparative mitochondrial genomic analyses between yeast species have revealed broad evolutionary changes in genome organization and architecture. A fine-scale view of recent evolutionary changes within S. cerevisiae has not been possible due to low numbers of complete mitochondrial sequences. To address challenges of sequencing AT-rich and repetitive mitochondrial DNAs (mtDNAs), we sequenced two divergent S. cerevisiae mtDNAs using a single-molecule sequencing platform (PacBio RS). Using de novo assemblies, we generated highly accurate complete mtDNA sequences. These mtDNA sequences were compared with 98 additional mtDNA sequences gathered from various published collections. Phylogenies based on mitochondrial coding sequences and intron profiles revealed that intraspecific diversity in mitochondrial genomes generally recapitulated the population structure of nuclear genomes. Analysis of intergenic sequence indicated a recent expansion of mobile elements in certain populations. Additionally, our analyses revealed that certain populations lacked introns previously believed conserved throughout the species, as well as the presence of introns never before reported in S. cerevisiae. Our results revealed that the extensive variation in S. cerevisiae mtDNAs is often population specific, thus offering a window into the recent evolutionary processes shaping these genomes. In addition, we offer an effective strategy for sequencing these challenging AT-rich mitochondrial genomes for small scale projects.

With growing of Semantic Web data, especially RDF data, managing large RDF dataset on a single machine does not scale well. Previous work has explored how to distribute RDF triples to multiple machines. However due to inefficient dataset partitioning used by these solutions, the performance of distributed store system is significantly affected. In this paper, we proposed a promising approach that utilized the graph nature of RDF datasets to minimize relations between partitions after dataset partitioning, and optimized system design based on it. As shown in our experiments, our approach can effectively reduce communication cost of query-processing messages, balance size of partitions compared with other approaches, and enhance parallelism through independent sub-querying.