Artificial intelligence (AI) is the hottest topic currently out there - no doubt about that. Neural networks in particular have seen a lot of attention and they will be used everywhere -self driving cars, predictions in finance and sales forecasts - everywhere and across all industries. To be successful in the working world of tomorrow we have to expose ourselves to this interesting topic - and from my personal experience - coding your own neural network is the best way to understand how they work. "From my personal experience I can tell you that companies will actively searching for you if you aquire some skills in the data science field. Diving into this topic can not only immensly improve your career opportunities but also your job satisfaction!"

The talk will cover various options for the most popular and widely used ML libraries, including MLeap, TF Serving and open standards such as PMML, PFA and the recently announced ONNX for Deep Learning. I will also introduce Aardpfark, initially covering Spark ML pipelines - as well as experimental work for exporting Spark ML pipelines to TensorFlow graphs for use with TF Serving.

Determining the localization of specific protein in human cells is important for understanding cellular functions and biological processes of underlying diseases. Among imaging techniques, high-throughput fluorescence microscopy imaging is an efficient biotechnology to stain the protein of interest in a cell. In this work, we present a novel classification model Twin U-Net (TUNet) for processing and classifying the belonging of protein in the Atlas images. Several notable Deep Learning models including GoogleNet and Resnet have been employed for comparison. Results have shown that our system obtaining competitive performance.

De-identification of electronic health records (EHR) is a vital step towards advancing health informatics research and maximising the use of available data. It is a two-step process where step one is the identification of protected health information (PHI), and step two is replacing such PHI with surrogates. Despite the recent advances in automatic de-identification of EHR, significant obstacles remain if the abundant health data available are to be used to the full potential. Accuracy in de-identification could be considered a necessary, but not sufficient condition for the use of EHR without individual patient consent. We present here a comprehensive review of the progress to date, both the impressive successes in achieving high accuracy and the significant risks and challenges that remain. To best of our knowledge, this is the first paper to present a complete picture of end-to-end automatic de-identification. We review 18 recently published automatic de-identification systems -designed to de-identify EHR in the form of free text- to show the advancements made in improving the overall accuracy of the system, and in identifying individual PHI. We argue that despite the improvements in accuracy there remain challenges in surrogate generation and replacements of identified PHIs, and the risks posed to patient protection and privacy.

It is currently unclear why adversarial examples are easy to construct for deep networks that are otherwise successful with respect to their training domain. However, it is suspected that these adversarial examples lie within some small perturbation from the network's decision boundaries or exist in low-density regions with respect to the training distribution. Using persistent homology, we find that deep networks effectively have ``holes'' in their activation graphs, making them blind to regions of the input space that can be exploited by adversarial examples. These holes are effectively dense in the input space, making it easy to find a perturbed image that can be misclassified. By studying the topology of network activation, we find global patterns in the form of activation subgraphs which can both reliably determine whether an example is adversarial and can recover the true category of the example well above chance, implying that semantic information about the input is embedded globally via the activation pattern in deep networks.

We investigate neural network training and generalization using the concept of stiffness. We measure how stiff a network is by looking at how a small gradient step on one example affects the loss on another example. In particular, we study how stiffness varies with 1) class membership, 2) distance between data points (in the input space as well as in latent spaces), 3) training iteration, and 4) learning rate. We empirically study the evolution of stiffness on MNIST, FASHION MNIST, CIFAR-10 and CIFAR-100 using fully-connected and convolutional neural networks. Our results demonstrate that stiffness is a useful concept for diagnosing and characterizing generalization. We observe that small learning rates lead to initial learning of more specific features that do not translate well to improvements on inputs from all classes, whereas high learning rates initially benefit all classes at once. We measure stiffness as a function of distance between data points and observe that higher learning rates induce positive correlation between changes in loss further apart, pointing towards a regularization effect of learning rate. When training on CIFAR-100, the stiffness matrix exhibits a coarse-grained behavior suggestive of the model's awareness of super-class membership.

High-dimensional representations often have a lower dimensional underlying structure. This is particularly the case in many decision making settings. For example, when the representation of actions is generated from a deep neural network, it is reasonable to expect a low-rank structure whereas conventional structures like sparsity are not valid anymore. Subspace recovery methods, such as Principle Component Analysis (PCA) can find the underlying low-rank structures in the feature space and reduce the complexity of the learning tasks. In this work, we propose Projected Stochastic Linear Bandit (PSLB), an algorithm for high dimensional stochastic linear bandits (SLB) when the representation of actions has an underlying low-dimensional subspace structure. PSLB deploys PCA based projection to iteratively find the low rank structure in SLBs. We show that deploying projection methods assures dimensionality reduction and results in a tighter regret upper bound that is in terms of the dimensionality of the subspace and its properties, rather than the dimensionality of the ambient space. We modify the image classification task into the SLB setting and empirically show that, when a pre-trained DNN provides the high dimensional feature representations, deploying PSLB results in significant reduction of regret and faster convergence to an accurate model compared to state-of-art algorithm.

We propose a novel automatic method for accurate segmentation of the prostate in T2-weighted magnetic resonance imaging (MRI). Our method is based on convolutional neural networks (CNNs). Because of the large variability in the shape, size, and appearance of the prostate and the scarcity of annotated training data, we suggest training two separate CNNs. A global CNN will determine a prostate bounding box, which is then resampled and sent to a local CNN for accurate delineation of the prostate boundary. This way, the local CNN can effectively learn to segment the fine details that distinguish the prostate from the surrounding tissue using the small amount of available training data. To fully exploit the training data, we synthesize additional data by deforming the training images and segmentations using a learned shape model. We apply the proposed method on the PROMISE12 challenge dataset and achieve state of the art results. Our proposed method generates accurate, smooth, and artifact-free segmentations. On the test images, we achieve an average Dice score of 90.6 with a small standard deviation of 2.2, which is superior to all previous methods. Our two-step segmentation approach and data augmentation strategy may be highly effective in segmentation of other organs from small amounts of annotated medical images.

We show that model compression can improve the population risk of a pre-trained model, by studying the tradeoff between the decrease in the generalization error and the increase in the empirical risk with model compression. We first prove that model compression reduces an information-theoretic bound on the generalization error; this allows for an interpretation of model compression as a regularization technique to avoid overfitting. We then characterize the increase in empirical risk with model compression using rate distortion theory. These results imply that the population risk could be improved by model compression if the decrease in generalization error exceeds the increase in empirical risk. We show through a linear regression example that such a decrease in population risk due to model compression is indeed possible. Our theoretical results further suggest that the Hessian-weighted $K$-means clustering compression approach can be improved by regularizing the distance between the clustering centers. We provide experiments with neural networks to support our theoretical assertions.

We propose a simple objective evaluation measure for explanations of a complex black-box machine learning model. While most such model explanations have largely been evaluated via qualitative measures, such as how humans might qualitatively perceive the explanations, it is vital to also consider objective measures such as the one we propose in this paper. Our evaluation measure that we naturally call sensitivity is simple: it characterizes how an explanation changes as we vary the test input, and depending on how we measure these changes, and how we vary the input, we arrive at different notions of sensitivity. We also provide a calculus for deriving sensitivity of complex explanations in terms of that for simpler explanations, which thus allows an easy computation of sensitivities for yet to be proposed explanations. One advantage of an objective evaluation measure is that we can optimize the explanation with respect to the measure: we show that (1) any given explanation can be simply modified to improve its sensitivity with just a modest deviation from the original explanation, and (2) gradient based explanations of an adversarially trained network are less sensitive. Perhaps surprisingly, our experiments show that explanations optimized to have lower sensitivity can be more faithful to the model predictions.