My Profile Photo

Personal Webpage of Max Horn

I'm a PhD Student in Machine Learning and Computational Biology at ETH Zürich and work on the development of deep learning methods for real world medical time series.
My interests include but are not limited to: Machine Learning for Healthcare, Probabilistic Modelling, Time Series Modelling and Interpretable Machine Learning.
Here I write about stuff I care about in the realm of science, programming and technology.

MLSS2020 Causality Lectures - a brief summary

The first week of the virtual Machine Learning Summer School in Tübingen is over and it is time to take a brief look back at the lessons learned and the experience made during this time. In the following I will briefly some of the insights I made during the first week.

Causality and Causal Inference lectures

The Causality lectures were held by Bernhard Schölkopf and Stefan Bauer. As a heads up, I am most definitely not an expert in this field and I am solely summarizing the most interesting points according to my personal opinion :).

Causality I

The first lecture by Prof. B. Schölkopf was a more general introduction to causality, the required terms needed to understand the literature, how to infer causal structure in smaller scale experiments with stochastic or deterministic relationships between observations and finally the implications of causality to semi-supervised learning.

Here I found the work on deriving causality for deterministic cases of particular interest (see this paper). In this work, the authors use an assumption termed the independence of input and mechanism to derive a causal inference rule which does not require any assumptions on noise. The assumption states, that $ p(C) $ (the probability distribution of the cause) and $p(E|C)$ (the probability distribution of the effect conditional on the cause) should be independent. This implies, that the distribution of the effect would then in some way be dependent on the function $f$ mapping from cause to effect and thus $Cov(log f’, p_C) = 0$ (encoding the independence assumption) and $Cov(log f^{-1’}, p_E) > 0$ (encoding the dependence between the function mapping from cause to effect and the distribution of the effect). This leads to the inference rule that $X \rightarrow Y$ if $\int \log | f’(x) | p(x) dx \leq \int \log | f^{-1}(y)| p(y) dy $. This can also be computed using empirical estimators for the slope of the function mapping between X and Y.

Finally, B. Schölkopf presented implications of causality in the domain of semi-supervised learning. In particular, if independence of input and mechanism is true, he shows that semi-supervised learning can theoretically not benefit learning in the causal direction. In other words when a model is trying to infer effect from cause (thus $p(E|C)$), additional data from the cause distribution $p(C)$ will not help the model learn due to $ p(C) $ being independent of $ p(E|C) $. In contrast, if learning is in the anti-causal direction, semi-supervised learning can be beneficial.

Causality II

The second talk by Stefan Bauer focused on how to infer Structural Causal Models and how causality can be integrated with Deep Learning approaches.

In the first part of his talk, Stefan showed bridges between causal inference and non-linear ICA and further shows that there will always be both a causal as well as anti-causal linear model if additive Gaussian noise is assumed. Afterwards, Stefan talks about time series models and how ODEs are perfect models for causal structure in the presence of time.

In the second half, the topic switches to describing a causal perspective on representation learning. Here the bridges between disentangles representations and causality become evident. This is due to the assumption, that a change in distribution of the data would arise from a sparse change in causal conditionals or causal mechanisms, thus leading to similar properties as desired in disentangled representations. Nevertheless, recent research shows that disentangled representations cannot be learned in a completely unsupervised way (see Locatello et al, ICML 2019), also leading to potential issues with the discovery of causal mechanisms. There is some hope though, as few labels seem to help with determining correct disentangled representations (see Locatello et al., ICML 2020).

Finally, Stefan talks about some exciting work on encoding causal structure into machine learning architectures. Here a decoder is designed to resemble the structure of a general Structural Causal Model and trained to match the observations of the training data. These Structural Causal Autoencoders (see Leeb et al., under review NeurIPS 2020) were shown to yield good performance in learning representations for generating images and transferring between similar tasks.

All in all some very exciting directions. I am looking forward to seeing the future development of Causality and Machine Learning.

comments powered by Disqus