Author Archives: Thijs van Ommen

Talk by Raghavendra Selvan (University of Copenhagen)

You are all cordially invited to the AMLab seminar on Tuesday May 9 at 16:00 in C3.163, where Raghavendra Selvan (University of Copenhagen) will give a talk titled “Segmenting Tree Structures with Probabilistic State-space Models and Bayesian Smoothing”. Afterwards there are the usual drinks and snacks!

Abstract: Segmenting tree structures is common in several image processing applications. In medical image analysis, reliable segmentations of airways, vessels, neurons and other tree structures can enable important clinical applications. We present a method for extracting tree structures comprising of elongated branches by performing linear Bayesian smoothing in a probabilistic state-space. We apply this method to segment airway trees, wherein, airway states are estimated using the RTS (Rauch-Tung-Striebel) smoother, starting from several automatically detected seed points from across the volume. The RTS smoother tracks airways from seed points, providing Gaussian density approximations of the state estimates. We use covariance of the marginal smoothed density for each airway branch to discriminate true and false positives. Preliminary evaluation shows that the presented method results in additional branches compared to base-line methods.

Talk by Zeynep Akata

You are all cordially invited to the AMLab seminar on Tuesday May 2 at 16:00 in C3.163, where Zeynep Akata will give a talk titled “Vision and Language for Multimodal Deep Learning”. Afterwards there are the usual drinks and snacks.

Scaling up visual category recognition to large numbers of classes remains challenging. A promising research direction is zero-shot learning, which does not require any training data to recognize new classes, but rather relies on some form of auxiliary information describing the new classes. Ultimately, this may allow to use textbook knowledge that humans employ to learn about new classes by transferring knowledge from classes they know well. We tackle the zero-shot learning problem by learning a compatibility function such that matching image-class embedding pairs are assigned a higher score than mismatching pairs; zero-shot classification proceeds by finding the label vector yielding the highest joint compatibility score. We propose and compare different class embeddings learned automatically from unlabeled text corpora and from expert annotated attributes. Attribute annotations performed by humans are not readily available for most classes. On the other hand, humans have a natural ability to determine distinguishing properties of unknown objects. We use detailed visual descriptions collected from naive users as side-information for zero-shot learning, to generate images from scratch and to generate visual explanations which justify a classification decision.

Akata, Label Embeddings for Image Classification, TPAMI 2016
Xian, Latent Embeddings for Zero-Shot Classification, CVPR 2016
Akata, Multi-Cue Zero-Shot Learning with Strong Supervision, CVPR 2016
Xian, Zero-Shot Learning: The Good, the Bad and the Ugly, CVPR’17
Karessli, Gaze Embeddings for Zero-Shot Learning, CVPR’17
Reed, Learning Deep Representations of Fine-Grained Visual Descriptions, CVPR 2016
Reed, Generative Adversarial Text to Image Synthesis, ICML 2016
Reed, Learning What and Where to Draw, NIPS 2016
Hendricks, Akata, Generating Visual Explanations, ECCV 2016

Talk by Philip Versteeg

You are all cordially invited to the AMLab seminar on Tuesday April 18 at 16:00 in C3.163, where Philip Versteeg will give a talk titled “Prediction and validation of causal effects in gene knockout experiments”. Afterwards there are the usual drinks and snacks!

Abstract: Causal questions abound across the empirical sciences, including basic biology, epidemiology, psychology and economics. Molecular biology is a particularly interesting area due to the ability in that field to perform interventions via diverse experimental techniques, such as measurements of gene expression levels under single gene knockouts. Such experiments are a key tool for dissecting causal regulatory relationships and provide an opportunity to validate causal discovery methods using real experimental data. In this talk, we provide results of an empirical assessment of several causal discovery algorithms using large-scale data from knockout experiments. We discuss several measures of performance defined using held-out observational and interventional data and find that while discovering system-wide causal structure remains difficult, especially when using only observational data, predicting the set of strongest causal effects is more feasible. We report that predictions of the strongest total causal effects based on a combination of interventional and observational data can be stable across performance measures and consistently outperform non-causal baselines.

Talk by Rianne van den Berg

You are all cordially invited to the AMLab seminar on Tuesday April 11 at 16:00 in C3.163, where Rianne van den Berg will give a talk titled “Graph convolutional networks as recommender systems”. Afterwards there are the usual drinks and snacks!

Abstract: Graph convolutional neural networks as introduced by Thomas Kipf and Max Welling have been used for link prediction and entity classification in undirected graphs, and more recently also in multi-relational directed graphs. In this talk I will present preliminary results on the application of relational graph convolutional networks to link prediction in recommender systems. I will discuss the possibility to merge interaction data between users and items with knowledge graphs and social networks, and show preliminary results on how this affects the performance in rating predictions.

Talk by Tineke Blom

You are all cordially invited to the AMLab seminar on Tuesday April 4 at 16:00 in C3.163, where Tineke Blom will give a talk titled “Causal Discovery in the Presence of Measurement Error”. Afterwards there are the usual drinks and snacks!

Abstract: Causal discovery algorithms can predict causal relationships based on several assumptions, which include the absence of measurement error. However, this assumption is most likely violated in practical applications, resulting in erroneous, irreproducible results. In this work, we examine the effect of different types of measurement error in a linear model of three variables, which is a minimal example of an identifiable causal relationship. We show that the presence of unknown measurement error makes it impossible to detect independences between the actual variables from the data using regular statistical testing and conventional thresholds for (in)dependence. We show that for limited measurement error, we can obtain consistent causal predictions by allowing for a small amount of dependence between (conditionally) independent variables. We illustrate our results in both simulated and real world protein-signaling data.

Talk by Frederick Eberhardt (Caltech)

You are all cordially invited to the AMLab seminar on Tuesday March 21 at 16:00 in C3.163, where Frederick Eberhardt (Caltech) will give a talk titled “Causal Macro Variables”. Afterwards there are the usual drinks and snacks!

Abstract: Standard methods of causal discovery take as input a statistical data set of measurements of well-defined causal variables. The goal is then to determine the causal relations among these variables. But how are these causal variables identified or constructed in the first place? Often we have sensor level data but assume that the relevant causal interactions occur at a higher scale of aggregation. Sometimes we only have aggregate measurements of causal interactions at a finer scale. I will present recent work on a framework and method for the construction and identification of causal macro-variables that ensures that the resulting causal variables have well-defined intervention distributions. We have applied this approach to large scale climate data, for which we were able to identify the macro-phenomenon of El Nino using an unsupervised method on micro-level sea surface temperature and wind measurements over the equatorial Pacific.

Talk by Taco Cohen

You are all cordially invited to the AMLab seminar on Tuesday March 14 at 16:00 in C3.163, where Taco Cohen will give a talk titled “Group Equivariant & Steerable CNNs”. Afterwards there are the usual drinks and snacks!

Abstract: Deep learning can be very effective, but typically requires large amounts of labelled data, which can be costly to collect. This is not only a major practical limitation to the applicability of deep learning, but also a fundamental barrier to AI: rapid learning is an essential part of intelligence.

In this talk I will present group equivariant networks, a natural generalization of convolutional networks that achieves improved statistical efficiency by exploiting symmetries like rotation and reflection. Instead of using convolutions, these networks use group equivariant convolutions. Group equivariant convolutions are easy to use, fast, and can be converted to standard convolutions after training. We show that simply replacing translational convolutions with group equivariant convolutions can improve image classification results. In the second part of the talk I will show how group equivariant nets can be scaled up to very large symmetry groups using steerable filters.

Talk by Karen Ullrich

You are all cordially invited to the AMLab seminar on Tuesday March 7 at 16:00 in C3.163, where Karen Ullrich will give a talk titled “Soft Weight-Sharing for Neural Network Compression”. Afterwards there are the usual drinks and snacks!

Abstract: The success of deep learning in numerous application domains created the desire to run and train them on mobile devices. This however, conflicts with their computationally, memory and energy intense nature, leading to a growing interest in compression. Recent work by Han et al. (2015a) propose a pipeline that involves retraining, pruning and quantization of neural network weights, obtaining state-of-the-art compression rates. In this paper, we show that competitive compression rates can be achieved by using a version of ”soft weight-sharing” (Nowlan & Hinton, 1992). Our method achieves both quantization and pruning in one simple (re-)training procedure. This point of view also exposes the relation between compression and the minimum description length (MDL) principle.

Talk by ChangYong Oh

You are all cordially invited to the AMLab seminar on Tuesday February 28 at 16:00 in C3.163, where ChangYong Oh will give a talk titled “High dimensional Bayesian Optimization”. Afterwards there are the usual drinks and snacks!

Abstract: Bayesian optimization has been successful in many hyper-parameter optimization problems and reinforcement learning problems. Still, there are many obstacles which prevent it from being extensively applied. Among many obstacles, we focused on the methods for high dimensional spaces. In order to resolve the difficulties of high dimensional Bayesian optimization problems, we devised a principled method to reduce the predictive variance of Gaussian process and other assistive methods for its successful application.
Firstly, brief explanation about general Bayesian optimization will be given. Secondly, I will explain the sources that make high dimensional problems harder, namely, ‘boundary effect’ and ‘hollow ball problem’. Thirdly, I will propose solutions to those problem, so-called, ‘variance reduction’ and ‘adaptive search region’.

Talk by Artem Grotov

You are all cordially invited to the AMLab seminar on Tuesday February 21 at 16:00 in C3.163, where Artem Grotov will give a talk titled “Deep Counterfactual Learning”. Afterwards there are the usual drinks and snacks!

Abstract: Deep learning is increasingly important for training interactive systems such as search engines and recommenders. They are applied to a broad range of tasks, including ranking, text similarity, and classification. Training neural network to perform classification requires a lot of labeled data. While collecting large supervised labeled data sets is expensive and sometimes impossible, for example for personalized tasks, there often is an abundance of logged data collected from user interactions with an existing system. This type of data is called logged bandit feedback and utilizing it is challenging because such data is noisy, biased and incomplete. We propose a learning method, Constrained Conterfactual Risk Minimisation (CCRM), based on counterfactual risk minimization of empirical Bernstein bound to tackle this problem and learn from logged bandit feedback. We evaluate CCRM on an image classification task. We find that CCRM performs well in practice and outperforms existing methods.