Physics Enhanced Microscopy for Virtual Fluorescence
In this work we demonstrate the principle of "Learned Sensing" applied to fluorescence microscopy. We jointly optimized the illumination of a sample with a deep neural network to predict fluourescent images from unlabeled samples. Our results demonstrate that by adding a cheap LED array to a conventional transmission microscope you can greatly improve performance on virtual fluorescence tasks, and likely In Silico labelling tasks more broadly.Publication:
Cooke, Colin L., et al. "Physics-enhanced machine learning for virtual fluorescence microscopy." arXiv preprint:arXiv:2004.04306 (2020).
Adaptive Microscope Illumination for Sample Classification
Here we demonstrate a new take on Learned Sensing by allowing the illumination of the sample to be governed by the sample. Looking at both simulated and experimental tasks we created a deep learning system that iteratively displayed different LED patterns on the sample, and integrated the coresponding images using an LSTM to make a classification.Publication:
Chaware, Amey*, Cooke, Colin L.*, et al. "Towards an intelligent microscope: adaptively learned illumination for optimal sample classification." IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020.
Movement Signal Processing
Worked with Professor Arash Arami at the University of Waterloo to link concepts within reinforcement learning and neuromechanics. Developed new techniques for policy exploration which are inspired by the Muscle Synergy Hypothesis.
Staged Training to Predict Sea Ice Concentration
Working with professor K. Andrea Scott at the University of Waterloo I developed deep learning training strategies to predict the sea ice concentration of Synaptive Aperature Radar images. We guided the training of CNNs by utilizing staged learning. We injected noise into the base image during training and gradually reduced it over several epochs. This led to the CNN learning feature based patterns prior to texture based and stopped the CNN from being stuck in a conceptual minima.
Cooke, Colin LV, and K. Andrea Scott. "Estimating sea ice concentration from sar: Training convolutional neural networks with passive microwave data." IEEE Transactions on Geoscience and Remote Sensing 57.7 (2019): 4735-4747.
Augmented Solar Analysis
While working for Heliolytincs Inc. I integrated the results of Aerial Image Segmentation into analysis workflows. I utilized known spatial information of the solar panels as well as adaptive pattern recognition to combine information from multiple aerial viewpoints and empowered analysts to focus on identification rather than localization.
Large Scale Neural-Simulations
Working with the Center for Theoretical Neuroscience at the University of Waterloo I experimented with possible optimization methods for biologically plausible nerual simulations. I built networks which learned how to acheive tasks ranging from simple multiplication to compression of images. In addition I led an investigation into optimizing very large models that could acheive complex tasks (like associative memory) by optimizing the sub-networks within them.
Segmentation of Aerial Imagery
While working for Heliolytics Inc. I pioneered the development of deep learning convolutional neural networks to segment aerial imagery. The models developed automate the analysis of solar panels allowing analysts to work faster and focus on edge cases rather than easily analyzed panels.