I’m a PhD student focusing on the application of machine learning to electron microscopy. This mainly means GPU-accelerated neural network training with TensorFlow. So far, I have written papers on
- Improving electron micrograph signal-to-noise, especially in low-dose micrographs: https://arxiv.org/abs/1807.11234. This is achieved using a large neural network based on Xception.
- Electron micrograph restoration and compression: https://arxiv.org/abs/1808.09916. This features 14 autoencoders, 15 kernels and 14 multilayer perceptrons. These can all be used for image restoration and the autoencoders have been trained for compression ratios up to 64:1.
- Neural networks that play with each other to reduce scanning transmission electron micrograph beam exposure and scan time: https://arxiv.org/abs/1905.13667. We are actively developing an experimental system.
- Adaptive learning rate clipping (ALRC) to limit backpropagated losses: https://arxiv.org/abs/1906.09060. ALRC complements existing learning algorithms to improve training stability, accelerate convergence and lower errors.
- Deep learning supersampled (DLSS) scanning transmission electron microscopy: https://arxiv.org/abs/1910.10467. This increases spatiotemporal resolution and decreases electron dose without scan system modification.
If you want to know more about what I’m doing, most of my work is on GitHub: https://github.com/Jeffrey-Ede
Datasets have been moved here.