Confirmation bias is the tendency to search for, interpret, favor, and recall information in a way that confirms or supports one's prior beliefs or values. People display this bias when they select information that supports their views, ignoring contrary information, or when they interpret ambiguous evidence as supporting their existing attitudes. The effect is strongest for desired outcomes, for emotionally charged issues, and for deeply entrenched beliefs. Confirmation bias cannot be eliminated, but it can be managed, for example, by education and training in critical thinking skills. Biased search for information, biased interpretation of this information, and biased memory recall, have been invoked to explain four specific effects:

attitude polarization (when a disagreement becomes more extreme even though the different parties are exposed to the same evidence)

belief perseverance (when beliefs persist after the evidence for them is shown to be false)

the irrational primacy

Exploring the coupled ocean and atmosphere system with a data science approach applied to observations from the Antarctic Circumnavigation Expedition

Andrea Baccarini, Irina Gorodetskaya, Christel Sybille Hassler, Sebastian Johannes Heinz-Josef Landwehr, Katherine Colby Leonard, Robin Lewis Modini, Fernando Perez Cruz, Julia Schmale, Jenny Thomas

The Southern Ocean is a critical component of Earth's climate system, but its remoteness makes it challenging to develop a holistic understanding of its processes from the small scale to the large scale. As a result, our knowledge of this vast region remains largely incomplete. The Antarctic Circumnavigation Expedition (ACE, austral summer 2016/2017) surveyed a large number of variables describing the state of the ocean and the atmosphere, the freshwater cycle, atmospheric chemistry, and ocean biogeochemistry and microbiology. This circumpolar cruise included visits to 12 remote islands, the marginal ice zone, and the Antarctic coast. Here, we use 111 of the observed variables to study the latitudinal gradients, seasonality, shorter-term variations, geographic setting of environmental processes, and interactions between them over the duration of 90 d. To reduce the dimensionality and complexity of the dataset and make the relations between variables interpretable we applied an unsupervised machine learning method, the sparse principal component analysis (sPCA), which describes environmental processes through 14 latent variables. To derive a robust statistical perspective on these processes and to estimate the uncertainty in the sPCA decomposition, we have developed a bootstrap approach. Our results provide a proof of concept that sPCA with uncertainty analysis is able to identify temporal patterns from diurnal to seasonal cycles, as well as geographical gradients and “hotspots” of interaction between environmental compartments. While confirming many well known processes, our analysis provides novel insights into the Southern Ocean water cycle (freshwater fluxes), trace gases (interplay between seasonality, sources, and sinks), and microbial communities (nutrient limitation and island mass effects at the largest scale ever reported). More specifically, we identify the important role of the oceanic circulations, frontal zones, and islands in shaping the nutrient availability that controls biological community composition and productivity; the fact that sea ice controls sea water salinity, dampens the wave field, and is associated with increased phytoplankton growth and net community productivity possibly due to iron fertilisation and reduced light limitation; and the clear regional patterns of aerosol characteristics that have emerged, stressing the role of the sea state, atmospheric chemical processing, and source processes near hotspots for the availability of cloud condensation nuclei and hence cloud formation. A set of key variables and their combinations, such as the difference between the air and sea surface temperature, atmospheric pressure, sea surface height, geostrophic currents, upper-ocean layer light intensity, surface wind speed and relative humidity played an important role in our analysis, highlighting the necessity for Earth system models to represent them adequately. In conclusion, our study highlights the use of sPCA to identify key ocean–atmosphere interactions across physical, chemical, and biological processes and their associated spatio-temporal scales. It thereby fills an important gap between simple correlation analyses and complex Earth system models. The sPCA processing code is available as open-access from the following link: https://renkulab.io/gitlab/ACE-ASAID/spca-decomposition (last access: 29 March 2021). As we show here, it can be used for an exploration of environmental data that is less prone to cognitive biases (and confirmation biases in particular) compared to traditional regression analysis that might be affected by the underlying research question.

2021

Bias at a Second Glance: A Deep Dive into Bias for German Educational Peer-Review Data Modeling

Vinitra Swamy

Natural Language Processing (NLP) has become increasingly utilized to provide adaptivity in educational applications. However, recent research has highlighted a variety of biases in pre-trained language models. While existing studies investigate bias in different domains, they are limited in addressing fine-grained analysis on educational and multilingual corpora. In this work, we analyze bias across text and through multiple architectures on a corpus of 9,165 German peer-reviews collected from university students over five years. Notably, our corpus includes labels such as helpfulness, quality, and critical aspect ratings from the peer-review recipient as well as demographic attributes. We conduct a Word Embedding Association Test (WEAT) analysis on (1) our collected corpus in connection with the clustered labels, (2) the most common pre-trained German language models (T5, BERT, and GPT-2) and GloVe embeddings, and (3) the language models after fine-tuning on our collected data-set. In contrast to our initial expectations, we found that our collected corpus does not reveal many biases in the co-occurrence analysis or in the GloVe embeddings. However, the pre-trained German language models find substantial conceptual, racial, and gender bias and have significant changes in bias across conceptual and racial axes during fine-tuning on the peer-review data. With our research, we aim to contribute to the fourth UN sustainability goal (quality education) with a novel dataset, an understanding of biases in natural language education data, and the potential harms of not counteracting biases in language models for educational tasks.

2022

Robustness and invariance properties of image classifiers

Apostolos Modas

Deep neural networks have achieved impressive results in many image classification tasks. However, since their performance is usually measured in controlled settings, it is important to ensure that their decisions remain correct when deployed in noisy environments. In fact, deep networks are not robust to a large variety of semantic-preserving image modifications, even to imperceptible image changes -- known as adversarial perturbations -- that can arbitrarily flip the prediction of a classifier. The poor robustness of image classifiers to small data distribution shifts raises serious concerns regarding their trustworthiness. To build reliable machine learning models, we must design principled methods to analyze and understand the mechanisms that shape robustness and invariance. This is exactly the focus of this thesis.First, we study the problem of computing sparse adversarial perturbations, and exploit the geometry of the decision boundaries of image classifiers for computing sparse perturbations very fast. We evaluate the robustness of deep networks to sparse adversarial perturbations in high-dimensional datasets, and reveal a qualitative correlation between the location of the perturbed pixels and the semantic features of the images. Such correlation suggests a deep connection between adversarial examples and the data features that image classifiers learn.To better understand this connection, we provide a geometric framework that connects the distance of data samples to the decision boundary, with the features existing in the data. We show that deep classifiers have a strong inductive bias towards invariance to non-discriminative features, and that adversarial training exploits this property to confer robustness. We demonstrate that the invariances of robust classifiers are useful in data-scarce domains, while the improved understanding of the data influence on the inductive bias of deep networks can be exploited to design more robust classifiers. Finally, we focus on the challenging problem of generalization to unforeseen corruptions of the data, and we propose a novel data augmentation scheme that relies on simple families of max-entropy image transformations to confer robustness to common corruptions. We analyze our method and demonstrate the importance of the mixing strategy on synthesizing corrupted images, and we reveal the robustness-accuracy trade-offs arising in the context of common corruptions. The controllable nature of our method permits to easily adapt it to other tasks and achieve robustness to distribution shifts in data-scarce applications.Overall, our results contribute to the understanding of the fundamental mechanisms of deep image classifiers, and pave the way for building more reliable machine learning systems that can be deployed in real-world environments.

EPFL2022