A5001453396- Electron and X-Ray Spectroscopy Techniques
- Advanced Electron Microscopy Techniques and Applications
- Machine Learning in Materials Science
- Advancements in Photolithography Techniques
- Genomic variations and chromosomal abnormalities
- Advanced X-ray Imaging Techniques
- Cell Image Analysis Techniques
- Photonic and Optical Devices
- Genomics and Rare Diseases
- Image Processing Techniques and Applications
- CRISPR and Genetic Engineering
- Model Reduction and Neural Networks
- Advanced Neural Network Applications
- Image and Signal Denoising Methods
- Digital Media Forensic Detection
- Domain Adaptation and Few-Shot Learning
- Sparse and Compressive Sensing Techniques
- Genomics and Phylogenetic Studies
Illumina (United States)
2023
University of Warwick
2019-2021
Personalized genome sequencing has revealed millions of genetic differences between individuals, but our understanding their clinical relevance remains largely incomplete. To systematically decipher the effects human variants, we obtained whole-genome data for 809 individuals from 233 primate species and identified 4.3 million common protein-altering variants with orthologs in humans. We show that these can be inferred to have nondeleterious humans based on presence at high allele...
Deep learning is transforming most areas of science and technology, including electron microscopy. This review paper offers a practical perspective aimed at developers with limited familiarity. For context, we popular applications deep in Afterwards, discuss hardware software needed to get started interface microscopes. We then neural network components, architectures, their optimization. Finally, future directions
Abstract Artificial neural network training with gradient descent can be destabilized by ‘bad batches’ high losses. This is often problematic for small batch sizes, order loss functions or unstably learning rates. To stabilize learning, we have developed adaptive rate clipping (ALRC) to limit backpropagated losses a number of standard deviations above their running means. ALRC designed complement existing algorithms: Our algorithm computationally inexpensive, applied any function size,...
Compressed sensing algorithms are used to decrease electron microscope scan time and beam exposure with minimal information loss. Following successful applications of deep learning compressed sensing, we have developed a two-stage multiscale generative adversarial neural network complete realistic 512$\times$512 scanning transmission micrographs from spiral, jittered gridlike, other partial scans. For spiral scans mean squared error based pre-training, this enables coverage be decreased by...
Personalized genome sequencing has revealed millions of genetic differences between individuals, but our understanding their clinical relevance remains largely incomplete. To systematically decipher the effects human variants, we obtained whole data for 809 individuals from 233 primate species, and identified 4.3 million common protein-altering variants with orthologs in human. We show that these can be inferred to have non-deleterious based on presence at high allele frequencies other...
Large, carefully partitioned datasets are essential to train neural networks and standardize performance benchmarks. As a result, we have set up new repositories make our electron microscopy available the wider community. There three main containing 19769 scanning transmission micrographs, 17266 98340 simulated exit wavefunctions, multiple variants of each dataset for different applications. To visualize image datasets, trained variational autoencoders encode data as 64-dimensional...
Abstract Compressed sensing can decrease scanning transmission electron microscopy dose and scan time with minimal information loss. Traditionally, sparse scans used in compressed sample a static set of probing locations. However, dynamic that adapt to specimens are expected be able match or surpass the performance as subset possible scans. Thus, we present prototype for contiguous system piecewise adapts paths they scanned. Sampling directions segments chosen by recurrent neural network...
Compressed sensing can increase resolution, and decrease electron dose scan time of microscope point-scan systems with minimal information loss. Building on a history successful deep learning applications in compressed sensing, we have developed two-stage multiscale generative adversarial network to supersample scanning transmission micrographs coverage reduced 1/16, 1/25, ..., 1/100 px. We propose novel non-adversarial policy train unified generator for multiple coverages introduce an...
Half of wavefunction information is undetected by conventional transmission electron microscopy (CTEM) as only the intensity, and not phase, an image recorded. Following successful applications deep learning to optical hologram phase recovery, we have developed neural networks recover phases from CTEM intensities for new datasets containing 98340 exit wavefunctions. Wavefunctions were simulated with clTEM multislice propagation 12789 materials Crystallography Open Database. Our can 224x224...
We present 14 autoencoders, 15 kernels and multilayer perceptrons for electron micrograph restoration compression. These have been trained transmission microscopy (TEM), scanning (STEM) both (TEM+STEM). TEM autoencoders 1$\times$, 4$\times$, 16$\times$ 64$\times$ compression, STEM 4$\times$ compression TEM+STEM 2$\times$, 8$\times$, 16$\times$, 32$\times$ Kernels to approximate the denoising effect of autoencoders. input sizes 3, 5, 7, 11 fitted TEM, TEM+STEM. with 1 hidden layer 5 7 2...
Compressed sensing can decrease scanning transmission electron microscopy dose and scan time with minimal information loss. Traditionally, sparse scans used in compressed sample a static set of probing locations. However, dynamic that adapt to specimens are expected be able match or surpass the performance as subset possible scans. Thus, we present prototype for contiguous system piecewise adapts paths they scanned. Sampling directions segments chosen by recurrent neural network based on...
This doctoral thesis covers some of my advances in electron microscopy with deep learning. Highlights include a comprehensive review learning microscopy; large new datasets for machine learning, dataset search engines based on variational autoencoders, and automatic data clustering by t-distributed stochastic neighbour embedding; adaptive rate clipping to stabilize learning; generative adversarial networks compressed sensing spiral, uniformly spaced other fixed sparse scan paths; recurrent...