PAC-Bayes Bounds for Bandit Problems: A Survey and Experimental Comparison.

PAC-Bayes has recently re-emerged as an effective theory with which one can derive principled learning algorithms with tight performance guarantees. However, applications of PAC-Bayes to bandit problems are relatively rare, which is a great misfortune. Many decision-making problems in healthcare, finance and natural sciences can be modelled as bandit problems. In many of these applications, principled algorithms with strong performance guarantees would be very much appreciated. This survey provides an overview of PAC-Bayes bounds for bandit problems and an experimental comparison of these bounds. On the one hand, we found that PAC-Bayes bounds are a useful tool for designing offline bandit algorithms with performance guarantees. In our experiments, a PAC-Bayesian offline contextual bandit algorithm was able to learn randomised neural network polices with competitive expected reward and non-vacuous performance guarantees. On the other hand, the PAC-Bayesian online bandit algorithms that we tested had loose cumulative regret bounds. We conclude by discussing some topics for future work on PAC-Bayesian bandit algorithms.

A Deterministic Approximation to Neural SDEs.

Neural Stochastic Differential Equations (NSDEs) model the drift and diffusion functions of a stochastic process as neural networks. While NSDEs are known to make accurate predictions, their uncertainty quantification properties have been remained unexplored so far. We report the empirical finding that obtaining well-calibrated uncertainty estimations from NSDEs is computationally prohibitive. As a remedy, we develop a computationally affordable deterministic scheme which accurately approximates the transition kernel, when dynamics is governed by a NSDE. Our method introduces a bidimensional moment matching algorithm: vertical along the neural net layers and horizontal along the time direction, which benefits from an original combination of effective approximations. Our deterministic approximation of the transition kernel is applicable to both training and prediction. We observe in multiple experiments that the uncertainty calibration quality of our method can be matched by Monte Carlo sampling only after introducing high computational cost. Thanks to the numerical stability of deterministic training, our method also improves prediction accuracy.

Towards brain-activity-controlled information retrieval: Decoding image relevance from MEG signals.

We hypothesize that brain activity can be used to control future information retrieval systems. To this end, we conducted a feasibility study on predicting the relevance of visual objects from brain activity. We analyze both magnetoencephalographic (MEG) and gaze signals from nine subjects who were viewing image collages, a subset of which was relevant to a predetermined task. We report three findings: i) the relevance of an image a subject looks at can be decoded from MEG signals with performance significantly better than chance, ii) fusion of gaze-based and MEG-based classifiers significantly improves the prediction performance compared to using either signal alone, and iii) non-linear classification of the MEG signals using Gaussian process classifiers outperforms linear classification. These findings break new ground for building brain-activity-based interactive image retrieval systems, as well as for systems utilizing feedback both from brain activity and eye movements.

Computer-aided diagnosis from weak supervision: a benchmarking study.

Supervised machine learning is a powerful tool frequently used in computer-aided diagnosis (CAD) applications. The bottleneck of this technique is its demand for fine grained expert annotations, which are tedious for medical image analysis applications. Furthermore, information is typically localized in diagnostic images, which makes representation of an entire image by a single feature set problematic. The multiple instance learning framework serves as a remedy to these two problems by allowing labels to be provided for groups of observations, called bags, and assuming the group label to be the maximum of the instance labels within the bag. This setup can effectively be applied to CAD by splitting a given diagnostic image into a Cartesian grid, treating each grid element (patch) as an instance by representing it with a feature set, and grouping instances belonging to the same image into a bag. We quantify the power of existing multiple instance learning methods by evaluating their performance on two distinct CAD applications: (i) Barrett's cancer diagnosis and (ii) diabetic retinopathy screening. In the experiments, mi-Graph appears as the best-performing method in bag-level prediction (i.e. diagnosis) for both of these applications that have drastically different visual characteristics. For instance-level prediction (i.e. disease localization), mi-SVM ranks as the most accurate method.

Event detection by feature unpredictability in phase-contrast videos of cell cultures.

In this work we propose a novel framework for generic event monitoring in live cell culture videos, built on the assumption that unpredictable observations should correspond to biological events. We use a small set of event-free data to train a multioutput multikernel Gaussian process model that operates as an event predictor by performing autoregression on a bank of heterogeneous features extracted from consecutive frames of a video sequence. We show that the prediction error of this model can be used as a probability measure of the presence of relevant events, that can enable users to perform further analysis or monitoring of large-scale non-annotated data. We validate our approach in two phase-contrast sequence data sets containing mitosis and apoptosis events: a new private dataset of human bone cancer (osteosarcoma) cells and a benchmark dataset of stem cells.

Empowering multiple instance histopathology cancer diagnosis by cell graphs.

We introduce a probabilistic classifier that combines multiple instance learning and relational learning. While multiple instance learning allows automated cancer diagnosis from only image-level annotations, relational learning allows exploiting changes in cell formations due to cancer. Our method extends Gaussian process multiple instance learning with a relational likelihood that brings improved diagnostic performance on two tissue microarray data sets (breast and Barrett's cancer) when similarity of cell layouts in different tissue regions is used as relational side information.

Automatic segmentation of colon glands using object-graphs.

Gland segmentation is an important step to automate the analysis of biopsies that contain glandular structures. However, this remains a challenging problem as the variation in staining, fixation, and sectioning procedures lead to a considerable amount of artifacts and variances in tissue sections, which may result in huge variances in gland appearances. In this work, we report a new approach for gland segmentation. This approach decomposes the tissue image into a set of primitive objects and segments glands making use of the organizational properties of these objects, which are quantified with the definition of object-graphs. As opposed to the previous literature, the proposed approach employs the object-based information for the gland segmentation problem, instead of using the pixel-based information alone. Working with the images of colon tissues, our experiments demonstrate that the proposed object-graph approach yields high segmentation accuracies for the training and test sets and significantly improves the segmentation performance of its pixel-based counterparts. The experiments also show that the object-based structure of the proposed approach provides more tolerance to artifacts and variances in tissues.