You are all cordially invited to the AMLab colloquium coming Tuesday May 10 at 16:00 in C3.163, where Stephan Bongers will give a talk titled “Marginalization and Reduction of Structural Causal Models”. Afterwards there are drinks and snacks!
Abstract: Structural Causal Models (SCMs), also known as (Non-Parametric) Structural Equation Models (NP-SEMs), are widely used for causal modelling purposes. One of their advantages is that they allow for cycles, i.e., causal feedback loops. In this work, we give a rigorous treatment of Structural Causal Models. Two different types of variables play a role in SCMs: “endogenous” variables and “exogenous” variables (also known as “disturbance terms” or “noise” variables). We define a marginalization operation (“latent projection”) on SCMs that effectively removes a subset of the endogenous variables from the model. This operation can be seen as projecting the description of a full system to the description of a subsystem. We show
that this operation preserves the causal semantics. We also show that in the linear case, the number of exogenous variables can be reduced so that only a single one-dimensional disturbance term is needed for each endogenous variable. This “reduction” can reduce the model complexity significantly and offers parsimonious representations for the linear case. We show under some suitable conditions this reduction is not possible in general.
You are all cordially invited to the AMLab colloquium coming Tuesday May 3 at 16:00 in C3.163, where Patrick Putzky will give a talk titled “Neural Networks for estimation in inverse problems”. Afterwards there are drinks and snacks!
Abstract: Many statistical problems arising in the natural sciences can be treated as an inverse problem: Measurements are transformed, subsampled or noisy observations of a quantity of interest. The main task is to infer the quantity of interest from the measurements.
Inverse problems are challenging because they are typically ill-posed. For example, if the number of variables in the quantity of interest exceeds the number observed variables, there is no unique solution to the inverse problem. To constrain the solution space the inverse problem is often phrased in terms of Bayes’ theorem. This allows to inject prior knowledge about the quantity of interest into the inference procedure. In practice, however, priors are often chosen to be overly simple (1) with respect to the complexity of the data, and (2) due to limitations in the inference procedure.
To overcome these limitations we propose an inference method which prevents the explicit notion of a prior. Instead, we suggest a neural network architecture which learns an inverse model for a given inference task. This approach has been frequently adressed before to solve problems such as image denoising, image deconvolution or image superresolution. However, the notion of the forward model has been mostly ignored in these approaches.
Our approach builds on previous neural network approaches for learning inverse models while explicitly making use of the forward model. The result is an iterative model which draws inspiration from gradient based inference methods. Our approach enables for learning a task specific inference model that has – compared to the traditional approach – the potential to (1) model complex data more reliably and (2) perform more efficiently in time critical tasks.
In the talk I will use the deconvolution problem in radio astronomy as a running example of an inverse problem, and on simulated data I will demonstrate how our approach compares to more traditional approaches. As a second example I will show some results for image superresolution.
You are all cordially invited to the AMLab colloquium today, April 12 at 16:00 in C3.163, where Philip Versteeg will give a talk titled “Causal inference and validation with micro-array data”. Afterwards there are drinks and snacks!
Abstract: In causality, one objective is to find cause and effect in a high dimensional setting. Often, identifiability of the true underlying graph is hard if not impossible when only observational data is available. A novel method, invariant causal prediction (ICP) is based on a relatively simple invariance principle. It is capable of exploiting a combination of observational and perturbed interventional experiments and provides confidence statements on the identified causal effects. In this talk, an application of ICP on high-dimensional micro-array experiments from the species Saccharomyces Cerevisiae is discussed. An attempt is made to validate the findings both internally and with external results.
You are all cordially invited to the next AMLab colloquium this Tuesday, February 23 at 16:00 in C3.163, where Joao Messias will give a talk titled “Variable-Order Markov Models for Sequence Prediction”. Afterwards there is the ‘borrel’ with drinks and snacks!
Abstract: The problem of learning how to predict future values of discrete stochastic sequences arises in many different domains. (Semi-)Markovian models with latent state representations, such as HMMs, have been widely used for this purpose, but their application can be difficult in domains with complex state or observation spaces. Recently, the first tractable approximations to the AIXI interpretation of Universal Reinforcement Learning have made use of Variable-Order Markov Models (VMMs) to approach this problem from a different perspective, by learning how to predict directly from histories of past observations, without requiring an explicit latent state representation, and in some cases, without knowing a priori how long these histories of observations should be.
In this talk, I will present an introduction to VMM methods for sequence prediction. I will discuss the relationship between VMM prediction and lossless compression; and present a review of the most well-known VMM methods. Finally, I will discuss future directions of research in this topic, in the context of reinforcement learning under partial observability.
You are all cordially invited to the next AMLab colloquium on Tuesday, February 16 at 16:00 in C3.163, where Mandar Chandorkar from the Multiscale Dynamics Group at CWI will give a talk titled “Space Weather Prediction using Gaussian Process (GP) Non Linear Auto-Regressive Models ”.
Abstract: Two models for predicting the Dst geo-magnetic time series are proposed and compared.
1. Non Linear Auto-regressive (GP-NAR)
2. Non Linear Auto-regressive GP with exogenous inputs (GP-NARX)
We present the results of extensive scale model testing experiments on the OMNI data set collected from the ACE satellite.
You are all cordially invited to the next AMLab colloquium on Tuesday, January 26 at 16:00 in C3.163, where Mijung Park will give a talk titled “Bayesian methodologies for efficient data analysis”.
Abstract: Machine learning and data science can greatly benefit from Bayesian methodologies, not only because they improve generalisation performance compared to point estimates that are prone to overfitting, but also they provide efficient and principled ways to solve a broad range of statistical problems. In this talk, I will describe several concrete examples where using Bayesian approaches greatly benefit in tackling problems occurring in many areas of science. These examples include (a) designing priors using domain knowledge for structurally sparse high-dimensional parameters with application to functional neuroimaging data and neural spike data; (b) Bayesian manifold learning that enables evaluating the quality of estimated latent manifold as well as learning the latent dimension from statistical evidence; and (c) approximate Bayesian computation (ABC) for models with intractable likelihoods, where we employ kernel mean embeddings to measure data similarities, which is an essential step in ABC.
You are all cordially invited to the next AMLab colloquium on Tuesday, January 19 at 16:00 in C3.163, where Deepak Geetha Viswanathan, will give a talk titled “Generalized parts-based models for unrectified images”.
Abstract: Parts-based detectors are a widely used class of models in object recognition. We propose a novel approach which takes lens distortion into account and generalizes parts-based detectors to unrectified images. Standard parts-based detectors are typically applied to unrectified images, which is sub-optimal, or to rectified images, which is time-consuming. By modifying the feature-extraction and the distance transform function to account for the distortion, we have developed a principled method to generalize parts-based detectors to unrectified images. We validate our approach on omni-directional images with a large amount of distortion, and empirically verify that our method outperforms the standard parts based detector trained on raw omni-directional images.
You are all cordially invited to the next AMLab colloquium on Tuesday, January 12 at 16:00 in C3.163, where Albert Huizing, principle scientist at TNO, will give a talk titled “Applying the art of deep learning to radar”.
Abstract: Deep learning techniques are currently successfully applied to object recognition in images, face recognition and speech recognition by major companies such as Google, Microsoft, Facebook, Apple, and Baidu. The application of deep learning to radar is at a more embryonic level than optical and acoustical sensors, partly because of the unknown benefits of radar in machine learning communities and partly because of the lack of large radar signal databases for training deep neural networks. The similarity of radar and optical images and radar and acoustical spectrograms, and the predicted growth of radar in commercial applications such as autonomous driving and gesture recognition, indicate that there is a need for investigating the application of deep learning in radar.
This talk will give a brief introduction of the principles of radar and provide a live demonstration of a small radar system. An overview of radar applications will show the benefits of radar when compared with other sensors such as cameras and microphones. In a case study, a model-based and a feature-based approach to human motion analysis with radar will be presented. The potential of deep learning for human motion analysis with radar will be discussed and concepts for fusing the signals of radars and cameras with multimodal deep neural networks will be explored.