By Claude Bajada, Simon M. Hofmann and Ilona Lipp
Edited by: Thomas Yeo and Lisa Nickerson
Machine learning, deep learning and artificial intelligence are terms that currently appear everywhere; in the media, in job adverts… and at neuroimaging conferences. Machine learning is often portrayed as a mystical black box that will either solve all our problems in the future or replace us in our jobs. In this blog post, we discuss what the term machine learning actually means, what methods it encompasses, and how these methods can be applied to brain imaging analysis. Doing this, we refer to the OHBM OnDemand material, which contains some great videos explaining machine learning methodology and we provide examples for how it has been used in a variety of applications. If you are curious about machine learning tools, but are not really sure whether you want to jump on the bandwagon, then we hope that this post is right for you and will help you get started.
What is machine learning?
Machine learning is a broad term that goes beyond deep learning and incorporates many other methods that are discussed in this post. Many of these methods you may already be familiar with and have heard about in the context of classical statistics, such as linear regression. While machine learning ‘is built on the foundations of statistics and has absorbed much of its philosophy and many of its techniques over the years’ (Max Welling, 2015), a main focus of machine learning lies in generalization, i.e. finding patterns in data by training a computational model such that it can predict unseen data of the same or similar nature. Here, a balance between “overfitting” (fitting a model that is too complex and will only work well for the data used to train it) and “underfitting” (fitting a model that is too simple and works poorly even with the data used to train it) needs to be found, aiming at high sensitivity and specificity when applied to new cases. This is generally attempted by splitting the data into various sets, training model parameters on one set, choosing the best model by evaluating it in another set, and testing it in yet another set. Sometimes, the term “statistical learning” is used for machine learning methods that have their foundation in statistics. Introduction to statistical learning and The Elements of statistical learning are two great textbooks introducing some important concepts.
What about Artificial Intelligence? Is that something different?
Artificial Intelligence is a high-level, conceptual term that describes the movement to emulate “natural intelligence” in machines. Machine learning can be thought of as one set of statistical tools that can make machines more “intelligent”.
What types of methods are there?
The most popular machine learning techniques applied to neuroimaging can be split into two general groups: supervised and unsupervised learning. Supervised learning requires labelled data (e.g., data that has been labelled, classified, or categorized), whilst unsupervised methods detect patterns in unlabelled data. Different machine learning methods also differ with regard to their complexity. Both types of machine learning approaches can range from fairly simple linear models to much more complicated, non-linear, algorithms. The more complex the models get, the more computational power is required.
Even though machine learning has been around for a long time, it has experienced a recent boom. In his OHBM OnDemand video, Vince Calhoun (6:30 min) explains why: it is not only because more and more data are available, but also because there has been an immense improvement in computational power (note that training neural networks can sometimes still take weeks!) and in better algorithms being developed and implemented in open source tools.
Below we will discuss some important methods that deploy both supervised and unsupervised learning. We will also discuss some approaches that are unique to neuroimaging (such as multivoxel pattern analysis).
As described by Christophe Phillips in the OHBM 2017 course on pattern recognition (4:34 min), the idea of supervised learning is to train a machine to find a mapping between observed data, such as your fMRI images, and an explanatory variable, which could be a disease label or a cognitive score. We can then find new, unlabeled subjects and predict the disease label or cognitive score.
Christophe further explains (7:10 min) that supervised machine learning problems can be further subdivided into discrete classification predictions and continuous, or regression, predictions. Either way, supervised learning (10:50 min) ultimately relies on a mapping function between input and target variables, the specification of the shape of this function and the optimization of its parameters. The following are some examples of various commonly used algorithms:
Most of you are familiar with linear regression as a classical statistical technique. However, this old staple has refashioned itself as a supervised learning technique. We can think of regression as a predictive technique that uses one (or many) features in order to predict a response as one continuous value (7:35 min). The main difference between using regression as a standard statistical tool and as a machine learning tool is that in machine learning we test the predictive power of the linear model on unseen data that did not contribute to the training of the model.
The idea behind logistic regression is, at its root, exactly the same as linear regression. The only difference is the function that is fitted to the data. While in linear regression we fit a line (or some generalisation of it in n dimensions – e.g, a plane or a hyperplane), in logistic regression we fit a logistic function. The logistic function is that “S-shaped” curve that often pops up in many biological sciences. The logistic function has the very nice property of being bounded (often these boundaries are set to 0 and 1) and hence can be used to express a probability. By having a cut-off, usually half way, we can use logistic regression to categorise our sample, for example into patients and controls.
Support Vector Machines (SVMs):
Support vector machines (SVM) are a type of classification algorithms, where the aim is to draw a decision (or classification) boundary between sets of data points, so as to maximize the “separation” (or margin) between the sets. While this sounds fairly straightforward, it is often the case that the data points are not easily separable by a line or plane, for instance, if two circles are embedded into one another. Kernel SVMs use “kernels” to transform the data into an alternative space where it might become much easier to separate the two instances. Christophe describes kernels and SVMs (from 17:00 min) in his introductory lecture. There are additional parameters such as regularisation parameters, gamma and the margin, that are important to define how well the line separates the training data. For a more generic discussion of SVMs, this medium post does a good job at explaining the basics.
Deep learning is one of the most talked about classes of machine learning algorithms and the one that most excites the public’s mind. Despite all the hype, deep learning models are often treated as a black box, since their input-output-mapping is both analytically and intuitively hard to grasp. In Vince Calhoun’s OHBM educational lecture on deep learning approaches applied to neuroimaging, he explains that the foundation of deep learning lies in artificial neural networks. In fact, despite experiencing a boom in popularity in recent years, neural network modelling dates back to the 1950s when there was a lot of interest in creating a mathematical model of a biological neuron (This paper by Hassabis et al. (2017) provides a stimulating discussion on the relationship between neuroscience and artificial intelligence). This neuronal model became known as a perceptron. The most basic type of network is the multilayer perceptron (MLP), with artificial neurons (perceptrons) organised in hierarchical layers. The input to the network is propagated layer-by-layer, first through activation-functions in each node, and then through connections (weights) to the successive layer. The “deep” part of deep learning refers to the number of multiple hidden layers, i.e. the layers between the input and output of the network. In recent years, computational advances have allowed the training of deeper and deeper networks. Some types of neural networks that Vince describes are Restricted Boltzmann Machines (7:10 min), deep belief networks (8:20 min), convolutional neural networks (16:35 min) and others.
As with other supervised learning algorithms, deep learning needs a training set and a test set. Furthermore, the more layers you have, the more (labeled) data and computational resources you usually need. In fact, deep learning increased in popularity once computing power increased to the point that deep networks were feasible, particularly after the availability of graphical processing units (GPUs), which are hardware chips originally developed for accelerated processing of digital videos and graphic rendering (3:10 min).
Multi-voxel pattern analysis (MVPA): A common application of ML in brain imaging?
In the classical analysis of structural and functional MRI, i.e. the application of a general linear model (GLM), each voxel is considered separately. Due to its linear equations, the approach is mathematically neat and tractable, however, this “massively univariate” approach disregards the interdependencies between multiple voxels (see Robert Cox’s talk about fMRI analysis methods at 4:16 min, Mike Pratt’s talk at 0:35 min). In light of dynamic brain processes that engage entire networks of the brain, the independence assumption of single voxels is controversial. In order to address this issue, a more recent class of statistical models, known as multi-voxel pattern analysis (MVPA), has been introduced to account for the joint contribution or ”combinatorial code“ of multiple voxels across the brain to the phenomenon of interest (see Janaina Mourão-Miranda’s talk at 6:08 min). That is, MVPA describes a class of pattern-recognition techniques, which are presented in Mike Pratt’s talk on MVPA (3:33 min), and in a session devoted to MVPA at OHMB 2017 (the corresponding videos can be found here).
MVPA draws from algorithmic strategies commonly used in machine learning. First, the data are split into a training set and test set. Then the classifier of choice (e.g., SVM) is trained on the former to discriminate various multi-voxel patterns corresponding to the experimental conditions, and validated on the latter. Validation is done by using the trained model to predict the conditions in the test set based on the multi-voxel input, which is often referred to as decoding (see Bertrand Thirion from 5:38 min, and Mike Pratt’s talk at 8:04 min). In decoding, we try to predict from multi-scale neural processes its representational content, such as percepts or cognitive states, mostly induced by experimental conditions (Pratt’s talk at 11:55 min). Classifiers can be linear or non-linear in nature, each having their own limitations. Linear classifiers (e.g., linear discriminant analysis, LDA) are considered easier to train and to interpret, however, their sensitivity depends on the individual contribution of each voxel in the observed pattern (see Jo Etzel’s talk at 18:00 min). Whereas non-linear classifiers (e.g., artificial neural networks, see Vince Calhoun’s talk) are able to find more complex relationships between patterns of voxels, they require training on large datasets.
The term MVPA was coined by Norman, Polyn, Detre, and Haxby (2006), who introduced it within the framework of fMRI analysis. However, considering a broader definition of the term, most of the methods that MVPA encompases are not restricted to fMRI and can be equally applied to structural imaging (e.g., Zhang et al., 2018; or Cole et al., 2017; and see James Cole’s talk at OHBM 2017).
In supervised learning, in addition to the input data (for example, fMRI images), we also need the ‘ground-truth’ output, which may be labels (e.g. healthy vs condition) or scores (some sort of cognitive or behavioural scores). However, frequently we either do not have appropriate labels, or the labels that we do have are unreliable, for example, in psychiatric imaging, as explained by Verena Kebets in her video. In this case, unsupervised machine learning methods open new doors.
In the neuroimaging community, the unsupervised machine learning technique of clustering is best known by its application to estimating brain parcellations. Brain parcellation is not a new problem and neither is it one that necessarily involves machine learning. All neuroimagers have heard of the 19th century neuroanatomist Korbinian Brodmann who labeled brain regions according to their cytoarchitecture -- the original brain mapping! As Simon Eickhoff explains in last year’s keynote, cytoarchitecture is not the only feature by which to parcellate the brain; there are others, such as receptor architecture, cortical myelin structure, and connectivity structure.
Unsupervised clustering methods are ideal for data with known differences based on features of interest where we want to automatically group brain regions according to these features. The simplest, and probably most widely used technique available, is k-means clustering. In neuroimaging, this is done by creating a feature vector per voxel in a region of interest, for example, structural or functional connectivity information. These voxels can now be thought of as points in an n-dimensional feature space. The k-means algorithm then attempts to maximize within-group similarity. Unfortunately, k-means clustering requires an a priori knowledge of the amount of groupings (k) one is interested in (although there are some iterative techniques to try to establish the number of k).
Other approaches to clustering, such as hierarchical clustering or spectral clustering, have the same basic idea of splitting up data (in this case brain voxels) into discrete groups, or parcels, but have slightly different assumptions or tricks. For example, hierarchical clustering assumes that the data have a hierarchical structure and so you could split the brain into two groups, each of which can be split into another two groups, until we reach the level of individual voxels. Or you could start from individual voxels and work your way up. On the other hand, spectral clustering has an additional step (the spectral transformation), which allows you to disregard weak similarity. Sarah Genon, in her educational course lecture, describes how to perform such analyses using diffusion MRI data.
Laplacian EigenMaps / Diffusion Embedding:
Sometimes you may not be interested in grouping voxels into a fixed number of parcels, but rather explore the relationship of voxels in a region of interest based on a feature of interest. In his educational talk, Daniel Margulies describes techniques that can be used to investigate the connectopies, or connectivity maps, of the brain. The initial approach is similar to the one described above, where you create a feature vector for every voxel in the brain. These features are then compared to each other using a measure of similarity to create a similarity, or affinity, matrix. This matrix is then decomposed and new vectors are obtained which describe the principal gradients of similarity across a region of interest, or indeed the whole brain. Daniel’s keynote describes how such types of analyses can be used to elucidate topographic principles of macroscale cortical connectivity.
Associative models, such as partial least squares (PLS) or canonical correlation analysis (CCA), are not exactly supervised or unsupervised. In supervised learning we generally have a multivariate input (e.g. brain images) and a univariate output (labels). In unsupervised learning, we only have one set of multivariate input data, such as the connectivity information for brain parcellation. In PLS or CCA, we want to discover relationships (associations) between two sets of multivariate inputs (e.g., brain images and behavioral/clinical scores).
As Janaina Mourao-Miranda explains in her video (2:25 min), psychiatric conditions often have unreliable labels. To deal with this, she uses associative models (e.g., PLS), trying to find a linear combination of neuroimaging predictors that are most strongly associated with a linear combination of multivariate clinical and behavioural data. This provides a data-driven way to generate summary labels that can possibly shed new light on clinical conditions.
It is possible to do significance testing on associative models to make inferences. Valeria Kebets describes (11:20 min) how to perform permutation tests in order to determine which components are significant, how to determine whether components are expressed differently across groups, and, finally, which variables drive the extracted components. In her video, Janaina also goes into the details about how her group applies a multiple hold-out validation framework in partial least squares analysis (16:50 min).
What do I need to consider when using machine learning tools for brain imaging analysis?
As explored in the previous paragraphs, machine learning techniques open many doors for brain imaging. They can help make predictions that depend on complex interactions, help find patterns in our data that we have been previously unaware of, and also automate time consuming manual tasks, such as segmentations (e.g. see Pim Moeskop’s video). However, there are also pitfalls that must be considered. First, the more complex and powerful machine learning techniques really need large datasets. In his video, Andrew Doyle (25:30 min) discusses how neuroimaging applications differ from classical image processing problems, with brain imaging data usually being very large and high-dimensional data, while sample sizes are comparatively small. For some applications (e.g. image segmentation or MVPA), smaller sample sizes may not be a big issue, but for others (such as patient classification) they may. A recent publication by Arbabshirani et al. (2018) explores the reason for why making individual predictions from brain imaging data is challenging. Another paper by Varoquaux (2018) focuses on the challenges with model cross validation on small sample sizes.
Of course, the noisier the data, the more data points are needed, and brain imaging data are renowned to be noisy. Additionally, if no reliable labels can be provided, the best supervised learning algorithms will not be able to succeed. Another problem, in particular with the more complex methods such as deep learning, is the challenge of assessing how biologically meaningful the resulting models are. Recent efforts have gone into better understanding and evaluating what is actually happening in the deep layers (e.g. watch Alex Binder’s video). However, resulting models may not teach us anything about biological or pathological mechanisms, and they may actually represent biases that exist in our training data, limiting their generalisability to other data. For example, this year’s replication award was awarded to a study that showed lack of generalisability of some published models.
Until these issues are fully resolved by the community, as individual researchers the best we can do is to understand the algorithms we are using and their limitations. That way we can choose the most suitable techniques, and rigorously apply them on suitable sample sizes and avoid overfitting. Luckily, there is a vast amount of online resources on machine learning techniques, including textbooks (e.g. Bishop, 2006), Andrew Ng’s famous Coursera courses on machine learning and deep learning, and online blogs and forums. Numerous papers from the MRI community provide overviews of machine learning tools for neuroimaging, or more specific examples, such as how machine learning is shaping cognitive neuroimaging, and how to use machine learning classifiers for fMRI data. OHBM’s OnDemand has an extensive archive of videos from education courses and talks on machine learning applications for neuroimaging that we’ve included in this article and, we also expect many exciting new educational and symposium talks on the use of machine learning techniques in brain imaging at this year’s OHBM in Rome, so watch out for those, too!
Professor David Kennedy is a Professor of Psychiatry at the University of Massachusetts. He was a key contributor to the development of functional MRI and diffusion MRI, working in MGH during the late 80s and 90s. His current work reflects his interests in Neuroinformatics and data sharing - indeed he is a founding editor of the journal Neuroinformatics. We found out about his experiences with OHBM, and some of the deep and lasting friendships he made along the way.
By Kirstie Whitaker
Open science means different things to different people. It includes open data, open source code, preprints, preregistrations, and open access publications. Getting started with open science practices can be overwhelming, and there is considerable variability in their adoption across the OHBM community. I sat down with Tibor Auer to learn about the survey he has developed to capture different attitudes towards open science practices in order to better support everyone in doing the best research they can.
Hi Tibor, let’s get started with an easy question: who are you?
I am a Research Fellow in MRI at the Department of Psychology, Royal Holloway University of London, where I facilitate neuroimaging by contributing to methods innovation, as well as training and education. Neurofeedback is one of my main research interests, as it offers the opportunity to follow neural development during the training process, thus satisfying interest in both theory and application. I received my PhD in clinical neuroscience and implemented various neuroimaging techniques in a clinical environment. Then, I focused on the implementation and the optimization of fMRI-based neurofeedback, and investigated assumptions and mechanisms underlying a neurofeedback training.
Why do you care about open science?
I am probably not the only person who has found a cool paper, and tried out its methods on my own data…. and fabulously failed! There are so many steps to reproducing a paper. If you are lucky you can find the corresponding author’s current e-mail address. They may bother to reply to your questions. The authors need to be able to locate the corresponding version of their workflow. It has to be documented well enough that you can decipher the code and adapt it to your data. Those are a lot of “ifs” and most of us aren’t that lucky. The process isn’t transparent.
Transparency, defined as unambiguous description of the data and the approaches, is not only beneficial to reproducibility but also to productivity in the first place. Automated pipelines, such as automatic analysis, allow quick exploration of the analysis space. Once you set up the workflow on pilot data, it can be applied on the real data right away. The pipeline’s documentation, based on standards such as the Neuroimaging Data Model, ensures the longevity of the project by making transitions smoother for current and future users.
What is this survey and what does it cover?
I am conducting this survey to investigate the knowledge and adoption of open research practices, including sharing of data and materials, study preregistration and related activities. It has been largely inspired by a recent survey in Cardiff. I want to know how people think about Open Science practices, their influence (positive or negative) and how the perception of Open Science might depend on experience with actual solutions and tools. Probably the most important aspect is what people see as the greatest barriers to the uptake of open research practices, and whether/how they can be ameliorated by (local and global) training and support. The awareness of challenges and solutions might vary across career levels, fields, and sectors (e.g. public and private), and I would like to be able to capture this variance, because we can only achieve change if we understand and resolve our differences.
Why did you want to create the survey?
Open Science is not just a (better) way to do science. One day, hopefully, we can omit the ‘open’, because all science will be done this way. Getting to that point cannot be an authoritarian process. It can only happen based on consensus and as a bottom-up initiative. We must understand what Open Science means for each of us, what encourages or discourages us to be engaged with it, what kind of support may be the most effective. The survey itself may raise awareness and give some hints and ideas by mentioning specific solutions in the questions.
Who do you hope will fill out the survey?
My guess is that many responses will be from methods experts who already appreciate the benefits of Open Science practices, but I would really like to also hear from the broader community of neuroimagers. Answers from people who do not use any Open Science practices are particularly valuable; especially if they tell us their reasons. It is important to know which aspects of Open Science have the biggest reach at the moment, and how they are perceived by a broad range of people in the OHBM community.
Early career researchers and senior scientists have different, sometimes even conflicting, priorities and motivations which may often explain the slow implementation of Open Science practices at an institutional level. Efficient resolution of these conflicts is possible only if we understand and harmonise the different incentives faced by these disparate groups.
How long will the survey be open for?
I have two stopping criteria for the survey: either we reach one thousand participants, or we get to 30 April 2019, whichever is earlier. I would like to receive responses from at least five hundred people but ideally one thousand!
I really appreciate the support of International Neuroinformatics Coordinating Facility (INCF) and OHBM in disseminating the survey to their members. It is important to note, however, that you do not have to be in either of these communities to respond. Everyone is welcome to submit their opinions.
What will you do with the results when you have them?
A summary of the results will be shared within the UK Network of Open Science Working Groups and other professional platforms, including the related Special Interest Groups of the OHBM and the INCF, to feed into the formulation and harmonisation of our Open Science strategy. The main aim of the survey is to capture different points of view, but I also hope it will prompt people to talk with each other, and think about and understand perceptions other than their own.
Every such survey has an educational angle, as well. For example, a recent survey on data management and sharing lets people know about several data handling approaches they might have not thought of before but they may try out after filling out the questionnaire. I hope this survey will encourage people to consider solutions unfamiliar to them and understand how they could benefit from additional tools.
Fantastic, I’m so looking forward to reading the results!
Thank you! Here’s the link again
By Amanpreet Badhwar and the Diversity and Gender Committee
The OHBM Diversity and Gender Committee is performing a series of interviews to better understand and address the issues of implicit and explicit biases in academia. The ultimate goal is to promote gender and geographic balance and create a more inclusive brain mapping community. Over the next months we will be interviewing social psychologists and social neuroscientists to get multiple perspectives on the topic. We start this series by interviewing Uta Frith.
Aman Badhwar (AB): How did you get into social psychology and neuroscience? Are there any personal experiences that motivated you?
Uta Frith (UF): When I left school for university, way back in the 1960s, in provincial Germany, I had never heard of neuroscience, and I wasn’t sure that psychology was a respectable subject to study. I more or less drifted into psychology and one reason was that I was curious to learn about myself and about other people. But that, I soon gathered, was considered the worst possible motivation for taking up psychology! Instead I could learn about memory, perception and attention. It sounded a bit dreary, but I persevered. I was not disappointed.
At the University of Saarbrücken, I discovered that there are millions of questions about the mind and the brain, but the most interesting ones came from some vivid case demonstrations in the psychiatric clinic. I therefore decided to take up training in clinical psychology and was lucky to get a place at the Institute of Psychiatry in London. But it was by chance that I met an autistic child right at the beginning of my training. This five-year-old boy was supremely uninterested in me, but very interested in bricks and puzzles. I was captivated by the strange contrast of abilities and disabilities. It made me determined to find out what might go on in the mind/brain of this little boy.
I never tired of doing research on autism and it taught me a lot about myself. As a child of my time, I had believed that it was my social environment that had turned me into a person keenly interested in other people. Now I had to entertain the notion that social interaction is possible only because of a built-in part of the human brain, and that if this part does not work, the result is autism.
Autism provided a way to probe the question ‘what makes human communication special’. In the 1980s a group of us tested the theory that a cognitive mechanism, we termed mentalising, enables us to attribute mental states to ourselves and others, and is crucial for reciprocal communication. It still feeds my sense of wonder that this cognitive mechanism is at the basis of human cooperation and reputation, as well as competition and deception.
AB: Women haven’t always received appropriate recognition for their work. How have you seen opportunities for women change? A change in climate? Is there a similar trend for women of colour?
UF: I grew up at a time of strong gender stereotypes and I was thoroughly imbued with them. My mother, born 1907, did not have a higher education, - it simply wasn’t even considered at the time. However, she managed to educate herself. At age eleven or so I begged my parents to send me to an academically challenging school that was then aimed at boys who would go on to University. The general view at the time was that girls should get a good all-round education at secondary school, then do something useful before getting married, such as becoming teachers or nurses. A few girls at my school were happily tolerated as oddballs or ‘bluestockings’. Since we were not conforming to the prevailing gender stereotype, we were having to shape our own role and station in life. I was diligent and ambitious, and I was sure that I would be able to take up any career I wanted. I don’t know where this confidence came from. Perhaps one reason was that my parents had confidence in me and always supported my choices.
I am very aware of the long history of the struggle for women’s rights. I am proud that in the last 100 years women in Western societies have won rights that previously were reserved for men. The fact that the inequality and inferiority of women has been seen as unsupportable, morally and economically, has changed the cultural climate. Now we need to work towards this change for every part of the world, not just for a few privileged parts.
However, there are still barriers even in privileged societies. For example, it is harder to enter an academic career if your social background leaves you ignorant or suspicious of the world of academia. And then there is still the glass ceiling! While women in the UK now make up almost a quarter of professors, when we take all universities and all disciplines together, women of colour make up a vanishingly small fraction. What is remarkable, is that women of colour have provided some of the most inspiring role models for women in science and technology in recent times – think of the success of the film Hidden Figures.
AB: What aspects related to gender differences within the work setting have not changed, but should?
UF: The first thing that comes to mind is the gender pay gap. Many organisations are now committed to close this gap, and they are deeply embarrassed if the gap is large. It would be a great injustice if women would not get the same pay, the same status, the same rewards for doing the same jobs as men. But this is a remarkably complicated issue, and it is not always easy to know what is the ‘same job’. I remember being surprised on discovering, when I became a professor, that my salary was lower than that of other professors. But then, I loved my job as a Medical Research Council scientist. It allowed me to dedicate myself entirely to research. I did not envy the administrative responsibilities or teaching duties that the higher paid professors had to do.
There are differences in our individual preferences and also abilities. I suppose these might correlate with gender differences to some extent. But individual differences trump gender differences. If you take only averages and forget about distributions and individual differences, then gender differences will emerge in some more or less relevant variables. Their importance depends on culture and context. I like to quote the example of a highly significant gender difference that exists in the ability to throw a ball. Women are far worse at this than men.
In situations where women compete with men, strengths and weaknesses associated with gender stereotypes often come to the fore. These can be quite ambiguous, sometimes harmless and sometimes treacherous, and therefore very suitable for sexist jokes. Q: “Is Google male or female?” A: “Female, because it doesn't let you finish a sentence before making a suggestion.” This makes me smile because it refers to a well-known stereotype but also contains a backhanded compliment – interrupting boring talk is surely better than meekly listening. But I can also see the exasperating side of making clever suggestions.
Of course, mostly, prejudices are not funny and can have serious consequences, for example in the selection of people given awards, or grants, or simply being invited to give talks. One thing we have learned from social cognitive neuroscience is that we cannot help identifying ourselves with a group, our ingroup. The reverse side of the coin is disrespecting an outgroup. The image we build of ourselves, and the confidence we have in ourselves feed on group alliance and group discrimination. None of us want to be seen as outsiders. We relish the approval of our ingroup and we perfectly understand if we are not exactly loved by the outgroup.
In the work setting there continues to be prejudices about gender differences that keep women down, or in their place, as some would have it. Early career researchers are necessarily at a stage of their life when they have to build up a family, and this almost always means a bigger burden on women than men. I am very glad to note that the younger generation of men are now taking up more of the burden. But this is still far from being the norm.
Is it possible to get rid of prejudices? Not really. I think that getting rid of one prejudice may well mean acquiring another. We have to use shortcuts when making fast decisions, or else we may never decide anything.
AB: Should there be more attention to microaggressions in the workplace, or at conferences? For example, ignoring or minimising ideas, claiming others’ ideas, assuming stereotypic roles. Have these reduced over time? Or increased with decrease of overt aggression?
UF: Wouldn’t it be great if we all became nicer people, more kind to each other and more tolerant? But that is a pipedream. We are shaped by evolution to be aggressive as well as benign, to be selfish as well as altruistic. I personally find it a satisfying side effect of getting older that competitiveness declines, and that with it the related emotions, such as pride, envy, revenge, or triumph, seem to decline as well. Sadly, I see less sign that my propensity for discriminating ingroups and outgroups declines with age.
Forms of aggression and selfish behaviour are expressed differently in different cultures, and there are subtle expectations that women should be less aggressive. They are also suspected of feeling offended more easily. Women can work against these stereotypes, and in particular not being afraid of giving and taking criticism. Academic life involves a lot of critique and judgment. It is hard to take justified critique and easy to take it for unjustified aggression. The main danger in my view is to feel offended when we shouldn’t be. Above all, we should not get drawn into retaliation. Empirical studies show that retaliation leads to escalation.
Hence my advice would be that monitoring should be less zealous when registering aggressions, including micro-ones, and more zealous when gauging the appropriate reactions to aggressions. Any monitoring is best done in diverse groups. With different backgrounds we have different perspectives, and we can see each other’s flaws much better than we can see our own.
AB: What tips would you give for young investigators, irrespective of their own gender to better understand gender biases, and to make changes?
UF: My main tip is to discuss and to argue freely any differences of opinion about whether some things are more suitable for men or for women. Take the opposite view and try the argument from the other side. Arguments are a good thing for scientists. We need diversity or else we can easily get stuck in a dead end.
I tend to think that complementarity of social roles and fairness to all should be celebrated over a relentless quest for equality. We are all different! Research needs different types of people, different skills, different outlooks and collaboration as well as competition. There are conflicts, and research tells us that transgressions need to be punished, or else cooperation collapses. But punishment is a dangerous thing. The danger is retaliation and escalation. So what to do? Rules of politeness have been set up for good reason and I recommend anyone to value them and not throw them aside as old fashioned. My advice goes further: try to be more than polite and try to be kind and forgiving! Of course, there will always be people who will do something that has rightly offended you or disadvantaged you in some way. But remember others may experience the same from you, even if you feel that you never gave cause for offence. Our self-bias makes this so.
AB: Are there any specific gender biases that earlier were advantageous for women, that are now gone?
UF: I can’t think of any.
AB: Thank you Uta for taking the time to provide such insight into the various topics. I really look forward to seeing you at OHBM 2019 in Rome!
Note: AB is also a member of the OHBM Diversity and Gender Committee.