Saturday, September 16, 2017
Friday, June 16, 2017
Since I wrote that blogpost I have found there are a lot of people out there who want to talk about retractions, the integrity of the scientific literature and the incentives researchers face around issues to do with scientific honesty.
Here a few of the things that have resulted from that original blogpost:
- David Duvenaud (who spotted the original bug in my code) created a presentation and a paper on the pitfalls of creating code for analysis, and sanity checks the analyst can use to avoid the same thing happening to them
- The story was picked up by Times Higher Education
- I was invited to tell my story at a symposium at the World Conference on Research Integrity 2017. The symposium was organised by Elizabeth Moylan, who with co-authors wrote a proposal for a new system of post-publication article alterations.
- After speaking at the symposium, the story of my retracted paper was covered by the founders of Retraction Watch in STAT and then by Science
|Speaking at the World Conference on Research Integrity|
Over time people have developed increasingly sophisticated theories of learning and education, and correspondingly teaching methods have changed and adapted. As a result, much is now known about what activities most promote student learning, and the differences between individuals in their learning techniques and strategies.
At the same time, computer scientists have developed increasingly powerful artificial intelligences. The creation of powerful computational methods for learning patterns, making predictions and understanding signals has drawn attention to a more mathematical understanding of how learning happens and can be facilitated.
Some of the parallels between these fields are obvious. For example, the development of artificial neural networks was driven by the analogy between these mathematical structures and the neuronal structure of the brain, and encouraged scientists to describe the brain from a computational perspective (e.g. in [Kovács, 1995]). However, the analogies between theories of learning in education and computer science are deeper than these surface resemblances, and go to the heart of what we consider useful information and knowledge, and what we mean by understanding.
In this report I will review elements of both the pedagogical and machine learning literature to draw attention to specific examples of what I consider to be direct analogues in these two fields, and how these analogies help organise our knowledge of the learning process and motivate approaches to student learning.
Learning to learn
When computer scientists first began creating an artificial intelligence, their first approach was to try to encode useful knowledge about the world directly in the machine, by explicitly inclusion in the computer’s programming. For example, in attempting to create a computer vision system that could recognise handwriting letters, the programmer would try to describe in computer code what an ‘A’ or a ‘B’ looked liked in terms that the computer could recognise in the images it received. However, this procedure generally proved dramatically ineffective. The sheer range of ways in which an ‘A’ can be written, the possible permutations on the basic design and the different angles and lighting that the computer could receive defeated the attempt to systematically describe the pattern in this top-down fashion.
The parallel here with the teacher-student relationship is very direct. A teacher is responsible, of course, for providing a great deal of information to a student. But the best teachers are more successful because they teach the students how to learn for the themselves, how to fit new examples into their existing understanding and how to seek the new information and examples they need. At the higher levels of tuition, encouraging and enabling this self-directed learning is essential. Anne Davis Toppins argues that within 30 minutes ‘I can convince most graduate students that they are self-directed learners’ [Toppins, 1987]. However, much as programmers initially tried to directly tell computers what they needed to know, before realising the greater efficiency of teaching them to learn for the themselves, so has the pedagogical approach taken a similar path [Gustafsson et al., 2011]:
'For some lecturers, thinking in terms of emphasising with and supporting the students’ learning and “teaching them to learn”, i.e. supporting them in their development of study skills, can constitute a new or different perspective. [...] Some teachers claim that since the students have studied for such a long time in other school situations, the higher education institution should not have to devote time to the learning procedure.'
In other words, there have been, and indeed still are many lecturers who view their role primarily in terms of transmitting information, rather than in developing the students’ abilities to think and learn for themselves.
In the modern teaching literature, much importance is placed on aiming for, and testing students conceptual knowledge. That is, students are expected to learn not simply a series of factual statements, or isolated results, but instead to incorporate their knowledge into higher level abstract concepts that they can use to understand unfamiliar situations, solve unseen problems and extrapolate their knowledge to new domains. The prevailing doctrine of constructive alignment [Biggs, 1999] that forms the basis for recommended teaching approaches in European countries under the Bologna process is designed to make sure that teaching methods, student activities and assessment assignments all align towards this goal of promoting and testing whether students understand the ‘big picture’.
According to a computer scientists view of knowledge and information, there is a very good reason why we should aim to promote such a concept-centred approach for students. Identifying unifying principles that tie knowledge together and understanding how apparently different fields may link together reduces the amount and the complexity of the information that a student or computer must store, access and process, and maximises the effectiveness of extrapolating to new domains.
Consider as a simple example the data shown in figure 1. How can this data be effectively stored? The simplest method would be the record each pair of (x, y) co-ordinates. Assuming we use a 1 byte per number (single-precision floating point accuracy), this will take us 20 bytes (10 x’s, 10 y’s). But visually we can immediately recognise an important pattern; the data clearly lie along a straight line. If we know the gradient of this line we can immediate translate any value of x into a value of y. Therefore we can reproduce the whole data set by specifying just 12 numbers – the 10 values of x, one value for the intercept and one value of the gradient. Therefore by understanding one big idea, one concept about the data, that they lie along a line, we have almost halved the effort of learning and storing that information. Furthermore, we can now extrapolate to any new slue of x, immediately knowing the correct corresponding value of y. If we had simply memorised the 10 pairs of co-ordinates we would have no way to do this. In the field on machine-learning this line of reasoning has been formalised into the principles of Minimum Message Length or Minimum Description Length, first proposed by Chris Wallace [Wallace and Boulton, 1968] and Jorma Rissanen [Rissanen, 1978] respectively. This states that the best model, or description of data set is the one which requires the least information to store. Modern texts on machine-learning theory focus heavily on the superiority of the simplest possible models that enable reconstruction of the necessary information and stress the connection to the well established principle of Occam’s Razor (e.g. [MacKay, 2003]). Applications of machine learning theory to animal behaviour have further suggested that animals apply the same principles to maximise the value of their limited processing and storage capabilities [Mann et al., 2011], so it is likely that humans also apply similar methods
Figure 1: By observing conceptual patterns in the data we can reduce the amount of memory needed to store it, whether on a machine or in a human mind. In this simple example identifying the linear relation between the X and Y co-ordinates (Y = 2X), we need to store only the X values, the intercept and the gradient, reducing the number of stored numbers from 20 to 12.
A common characteristic of high-achieving students is a strategic approach to learning. They have a good overview of what they need to learn to achieve their life goals. They set realistic but challenging learning goals for themselves to the end of learning this material. And they actively seek out information from teachers, reading materials and other sources to aid their learning. Whether their goals are intrinsic (interest in the subject, desire for knowledge) or extrinsic (obtaining a degree, getting a job), this strategic approach to learning systematically produces better outcomes than passively receiving whatever information is offered.
Analogously, in the field of machine learning, recent developments have tended more and more towards ideas termed ‘active learning’ [Settles, 2010]. The previous paradigm of simply offering many examples to the computer to learn from and then assessing or using the results of that process has been overturned. Instead, the programmer/mathematician devises a strategy for the computer to seek out new examples, based on what it wants to achieve (e.g. identifying written letters successfully) and what it currently knows. For example, if the computer has a good idea how to recognise an ‘A’, but frequently confuses a ‘U’ and a ‘V’, it will seek out or request more examples of these letters so that it can improve its knowledge. This way it does not waste time learning redundant material, but maximises the result of its effort by focusing on the most rewarding areas.
Likewise a high-performing student will focus their attentions on areas where they are weak and/or particularly crucial concepts that provide a pivot for understanding. They will ask their teachers for more feedback on their efforts in these areas, spend more time on mastering them and prioritise them ahead of areas of less importance or that are already understood. Mckeachie’s Teaching Tips [McKeachie and Svinicki, 2013] devotes a chapter to the importance encouraging strategic and self-regulated learning. One of their descriptions of a strategic learner states:
‘Strategic learners know when they understand new information and, perhaps more important, when they do not. When they encounter problems studying or learning, they use help-seeking strategies’.
This emphasis on the importance of know where understanding is lacking and the resultant help- seeking strategy perfectly aligns with what information theory tells us is the optimal way to gain useful knowledge.
Mckeachie’s Teaching Tips [McKeachie and Svinicki, 2013] also focuses on the importance of student learning goals. My own research in the field of active-learning corroborate this view, demonstrating that even when a learner has a good learning strategy, the success of that strategy depends intimately on the goals that the learner sets themselves. Indeed, without a suitable goal the learner is unable to define a useful strategy [Garnett et al., 2012]. Thus, in order to develop students strategic learning skills, it is essential first to help them define, and identify what their individual goals are. A student for whom this is an essential course, but who is otherwise uninterested, may be best helped by helping them to clarify what they wish to achieve (a certain final grade for instance), and then working with them to establish what strategy will most likely allow them to reach that outcome. A student with greater intrinsic motivation for the course may need help setting specific staged learning goals that enable a learning strategy. The teacher’s experience in understanding the most effective path through the material would therefore be essential in establishing effective goals that the student can then apply a strategy to achieve.
While student and machine learning are clearly not direct parallels of each other (could one imagine a machine equivalent for tiredness, or skipping class to watch TV?), the analogies that do exist be- tween the two help us to understand why certain approaches to student learning are more successful than others, via the large body of technical knowledge that exists regarding how machines can be taught. In this report I have analysed a selection of those analogies, aiming to draw conclusions about how students should be taught.
In particular, a common theme of modern pedagogical approaches is to move from information transfer to a student directed learning approach. In a sense, computer scientists have been down this path already, switching from a programmer-led to a computer-led learning approach that has resulted in far superior learning outcomes. This should motivate and support the equivalent transition in student learning
In teaching computers how to think and learn, we have also needed to help them establish goals and strategies for learning, and this is now the forefront of machine learning research. The dramatic improvement in computer learning outcomes when well-developed strategies are employed should remind us that it is the manner in which the student approaches new information and requests help and feedback that matter at least as much as the amount of information they are presented with. Such knowledge demands that we devote time to monitoring and developing students learning strategies and discussing what they hope to achieve via our courses.
Students, like all of us, are presented with a great deal more information than they can easily process and digest. If computer science in the 21st century has taught us anything, it is the importance of identifying general patterns in the vast body of information we are now exposed to via the media, the Internet and other sources. Without relatively simple general principles, information can easily become overwhelming. That the same principle applies in student learning should not surprise us. How is a student to retain all the information we attempt to transfer to them without organising it into general principles rather than a huge array of specific cases? The content of any course therefore should revolve as much around this organisational structure as the raw information itself, demanding generalised understanding rather than specific regurgitation. Thankfully this is the direction modern pedagogy is taking, with such concepts of constructive alignment and the SOLO taxonomy.
[Garnett et al., 2012] Garnett, R., Krishnamurthy, Y., Xiong, X., Schneider, J., and Mann, R. (2012). Bayesian optimal active search and surveying. In Proceedings of the International Con- ference of Machine Learning.
[Gustafsson et al., 2011] Gustafsson, C., Fransson, G., Morberg, ̊A., and Nordqvist, I. (2011). Teaching and learning in higher education: challenges and possibilities.
[Kovács, 1995] Kovács, I. (1995). Maturational windows and adult cortical plasticity, volume 24. Westview Press.
[MacKay, 2003] MacKay, D. J. C. (2003). Information Theory, Inference and Learning Algorithms. Cambridge: Cambridge University Press.
[Mann et al., 2011] Mann, R., Freeman, R., Osborne, M., Garnett, R., Armstrong, C., Meade, J., Biro, D., Guilford, T., and Roberts, S. (2011). Objectively identifying landmark use and predicting flight trajectories of the homing pigeon using gaussian processes. Journal of The Royal Society Interface, 8(55):210–219.
[McKeachie and Svinicki, 2013] McKeachie, W. and Svinicki, M. (2013). McKeachie’s teaching tips. Cengage Learning.
[Rissanen, 1978] Rissanen, J. (1978). Modeling by shortest data description. Automatica, 14(5):465–471.
[Settles, 2010] Settles, B. (2010). Active learning literature survey. University of Wisconsin, Madison, 52:55–66.
[Simon, 2013] Simon, P. (2013). Too Big to Ignore: The Business Case for Big Data. John Wiley & Sons.
[Toppins, 1987] Toppins, A. D. (1987). Teaching students to teach themselves. College Teaching, 35(3):95–99.
[Wallace and Boulton, 1968] Wallace, C. S. and Boulton, D. M. (1968). An information measure for classification. The Computer Journal, 11(2):185–194.
Wednesday, June 7, 2017
See my recent article in The Conversation for some reasons why betting markets may have been performing so badly in predicting elections and referenda in recent years, or read the original research here
Tomorrow is the polling day in the UK General Election 2017 (make sure you vote!). Today's news will be full of the latest opinion poll numbers, and pundits making predictions. Increasingly people are also looking to betting and prediction markets to get an idea of what is likely to happen as well. Both opinion polls and betting markets have made some very significant errors in recent years. Before the Brexit referendum I did an analysis of what bets on Betfair were telling us about the predicted vote share for Leave/Remain. Punters got that one wrong, just like the election of Trump in the USA, while polls were more accurate in predicting tight races.
Before we go to the polls tomorrow, lets compare what opinion polls and betting markets are telling us, so we can evaluate which is more accurate on this occasion. I'll focus simply on raw vote share for the two main parties (ignoring constituency effects), and I'll use the Financial Times poll-of-polls as a benchmark for the opinion polls and Betfair's vote share markets for betting markets.
First the opinion polls: https://ig.ft.com/elections/uk/2017/polls/
This gives a central forecast of Conservatives on 43%, Labour on 37%
To calculate the predicted vote share from Betfair I'll be repeating the analysis I did here (see previous post for R code), fitting a beta-distribution to the vote share divisions given on the market. I've taken screenshots of the Conservative and Labour markets, as these will no doubt change after I post this:
Performing the analysis to get the predicted vote share gives the following results:
This puts the Conservatives on 44% and Labour on 34% - almost identical for the Conservatives as the opinion poll, but somewhat lower for Labour.
Labour have recently surged in the polls from a very low position. It seems that the betting markets don't fully trust this. Come tomorrow night we'll have a good idea which of the polls or the market has been more accurate.
Tuesday, May 2, 2017
There’s a problem with the wisdom of crowds.
Market economies and democracies rely on the idea that whole populations know more about what is best for them than a small elite group. This knowledge is potentially so powerful it can even predict the future through stock markets, betting exchanges and special investment vehicles called prediction markets.
These markets allow people to trade “shares” in possible future outcomes, such as the winner of upcoming elections. Anyone with new information about the future has a financial incentive to spread it by buying these shares. Prediction markets now routinely inform bookmakers odds and are quoted in news coverage of elections alongside more traditional opinion polls.
But prediction markets are having a crisis of confidence in the abilities of the crowd. They have been systematically wrong about a series of high profile political decisions, including the UK general election of 2015, the Brexit referendum and the US presidential election of 2016.
We shouldn’t expect perfect accuracy on every occasion, just as we know opinion polls are often flawed. But to be wrong so consistently about such prominent events points to possible flaws in the assumptions we make about crowd intelligence. For example, people don’t always act on the information they have and so it might never become part of the crowd’s decision. The dynamics of crowds and markets might also stop people from paying attention to some sources of information at all.
However, there might be a way forward. My colleagues and I have come up with a model that overcomes this problem by giving people a incentive to seek out new sources of information, and an extra reason to share it.
An important question for markets is “where do individuals get their information?” Research shows that our opinions and activities very often match those of our peers. We also tend to look for information in the most obvious places, in line with everyone else.
To give an example, if you look around on any public transport in the City of London you’ll probably see people holding copies of the Financial Times. This is a problem because if everyone has the same information, the crowd is no smarter than a single individual. Studies show that having a diverse collection of opinions, especially including minority views, is crucial for creating a smart group.
So why do we tend to narrow the sources of our opinions? One reason is because we have an innate desire to imitate our peers, to behave in ways that are safe and acceptable within our community. But it may also be because of a rational, profit-seeking motivation.
We studied how theoretical profit-motivated people behave when faced with the types of rewards seen in market-like situations. To do this, we created a computer simulation of a prediction market, where people received a reward for making correct predictions. Rewards were larger when fewer people guessed the right answer, just like in a prediction market or a betting exchange.
The reward an individual received was a fixed amount divided by the number of other people who made a correct prediction. This was supposed to give people an incentive to look for right answers that other people wouldn’t find. But we found that people still gravitated towards a very small subset of the available information – just like London bankers with their copies of the Financial Times.
The more complex the situation was, the smaller the percentage of available information people actually used. The problem was that the more niche, unused information, though it might be useful to the group, was so rarely useful to the individual that possessed it that there was no incentive for them to seek it out.
New reward system
To counter this, we created a theoretical new prediction market system, where people would only be rewarded if they expressed accurate views but were also in the minority. For example, if someone predicted that Donald Trump would win the US election, against the consensus view, they would have received a reward once the result was known. Conversely, if most people accurately predict the Conservative Party will win the upcoming UK election then they wouldn’t receive any reward.
We found that this “minority reward” system, which explicitly favours those who go against popular opinion if they turn out to be correct, produced much more accurate collective decisions. This was especially the case when the situations were complex, influenced by many factors.
Intuitively, this makes sense. If your opinion supports the existing popular view, you can’t change whether the group will be correct or not. In our model, people have an incentive to go hunting for more esoteric sources of information about possible future outcomes. For example, rather than reading the Financial Times, they might follow obscure blogs, or read local newspapers looking for information on companies in the area.
They know that only by finding information that very few have access to will they have a chance to correctly go against the prevailing wisdom. This encourages the whole group to bring together a much wider set of information, leading to more accurate collective decisions.
Our results are so far confined to a theoretical model, but they give us an insight into why current forms of prediction markets may be prone to failure, and how we might try to improve them in future. We hope that these insights will be used to create more accurate prediction markets, as we could all benefit from better collective foresight.
Better predictions and collective decision making could help society decide which political ideas will or won’t work. Improving the ability of stock markets to predict which companies and ideas will do well could improve the return on investment and generate greater economic growth. Even academia is a large-scale exercise in collective wisdom. If changing the way that researchers are rewarded can improve the wisdom of this crowd, it could lead to more important scientific discoveries.
Tuesday, December 20, 2016
Each of these aspects of data-driven healthcare has promise, but also problems. It is very difficult to reliably associate a disease or drug efficacy with a small number of testable gene alleles, and very easy to identify false positive gene associations. Routinely collected data is very difficult to make reliable inferences from in terms of cause and effect, because treatments are not randomly assigned to patients. Sophisticated analytics do not stop you needing to think about how your data was collected. Lifestyle optimisation via smartphones probably owes more to Silicon Valley's ideal of the hyper-optimised individual and a corporate desire for ever more personal data than any real health benefits beyond an increased motivation to exercise.
However, there are easy wins to be had from data. These are in prediction of future events that involve no medical intervention. It is difficult to predict how a drug will affect a patient, because you need to infer the drug's effect against a background of other potential causes. But it is much easier to tell if a patient arriving at the hospital for a specific operation will need to stay overnight; simply look at whether similar patients undergoing similar operations have done so. If this sounds exceptionally simple, that's because it is. However, the gains could be great. Hospitals routinely have to keep expensive beds available to deal with emergencies, or cancel planned operations to deal with unexpected bed shortages. A reliable system to estimate the length of patient stay after an operation with some accuracy would reduce the need for these expensive, time consuming and inconveniencing issues. On the ground staff already have a good sense for which patients will need to stay longer than others. However, in the maelstrom of an NHS hospital, anything that can help to systematise and automate the making and use of these estimates will reduce pressures on staff.
Exploring this possibility, we performed an analysis of data the NHS routinely collects for patients and procedures, such as age, year, day and surgery duration (see figure below), and used this to predict stay duration. Our results showed that a substantial portion of the variability in stay duration could be predicted from these data, which would translate to a significant saving for the NHS if generally applied and combined with current estimates of stay given by experts on the ground from their past experience. Note, importantly, we are not suggesting any intervention on the individual as a result of this analysis. For instance we make no judgement on whether the variation by day indicates anything important about treatment, only that this helps planners to know whats likely to come up next. This work is not about whether the NHS should operate a full weekend service!
As with numerical weather forecasts, we envisage this supplementing and supporting existing human expert judgement, rather than replacing it - there are clearly facets of the patient that we cannot capture in a simple data analysis. This provides a minimal cost use of existing data, with little or no complicating causal issues, that could save the NHS money on a daily basis. The size of the NHS means that small gains can be amplified on a national scale, while NHS data provides an enormous potential resource. It may be in these unglamorous aspects of healthcare provision that data analytics has immediate potential.
Monday, December 5, 2016
As someone who has dabbled in criminology research, I was pretty disturbed by the paper from an ethical standpoint. I think this subject, even if it is declared fair game for research, ought to be approached with the utmost caution. The findings simply appeal too strongly to some of our more base instincts, and to historically dangerous ideas, to be treated casually. The sparsity of information about the data is troubling, and I personally find the idea of publishing photos of "criminals and non-criminals" in a freely-available academic paper to be extremely unsettling (I'm not going to reproduce them here). The paper contains no information on any ethical procedures followed.
Aside from these issues, I was also disappointed from a statistical perspective, and in a way that is becoming increasingly common in applications of machine-learning. The authors of this paper appear not to have considered any possible issues with the causality of what they are inferring. I have no reason to doubt that the facial patterns they found in the "criminal" photos are distinct in some way from those in the "non-criminal" set. That is, I believe they can, given a photo, with some accuracy predict which set it belongs to. However, they give no consideration to any possible causal explanation for why these individuals ended up in these two sets, beyond the implied idea that some individuals are simply born to be criminals and have faces to match.
Is it not possible, for example, that those involved in law enforcement are biased against individuals who look a certain way? Of course it is. Its not like there isn't research on exactly this question. Imagine what would happen if you conducted this research in western societies: do you doubt that the distinctive facial features of minority communities would be inferred as criminal, simply because of well-documented police and judicial bias against these individuals? In fact, you need not imagine, this already happens: machine-learning software analyses prisoners risk of reoffending, and entirely unsurprisingly attributes higher risk to black offenders, even though race is not explicitly included as a factor.
If this subject matter was less troublesome, I would support the publication of such results as long as the authors presented the findings as suggesting avenues for future, more careful controlled studies. However, in this case the authors resolutely do not take this approach. Instead, they conclude that their work definitively demonstrates the link between criminality and facial features:
"We are the first to study automated face-induced inferenceThis paper remains un-reviewed, and let us hope it does not get a stamp of approval by a reputable journal. However, it highlights a problem with the recent fascination with machine-learning methods. Partly because of the apparent sophistication of these methods, and partly because many in the field are originally computer scientists, physicists or engineers, rather than statisticians, there has been a reluctance to engage with statistical rigour and questions of causality. With many researchers hoping to be picked up by Google, Facebook or Amazon, the focus has been on predictive accuracy, and on computational efficiency in the face of overwhelming data. Some have even declared that the scientific method is dead now that we have Big Data. As Katherine Bailey has said: "Being proficient in the use of machine learning algorithms such as neural networks, a skill that’s in such incredibly high demand these days, must feel to some people almost god-like ".
on criminality. By extensive experiments and vigorous
cross validations, we have demonstrated that via supervised
machine learning, data-driven face classifiers are able
to make reliable inference on criminality. Furthermore, we
have discovered that a law of normality for faces of noncriminals.
After controlled for race, gender and age, the
general law-biding public have facial appearances that vary
in a significantly lesser degree than criminals."
This is dangerous nonsense, as the claim to infer criminality from facial features shows. It is true that Big Data gives us many new opportunities. In some cases, accurate prediction is all we need, and as we have argued in a recent paper, prediction is easy, cheap and unproblematic compared to causal inference. Where simple predictions can help, we should go ahead. We absolutely should be bringing the methods and insights of machine-learning into the mainstream of statistics (this is a large part of what I try to do in my research). Neil Lawrence has said that Neural Networks are "punk statistics", and by God statistics could do with a few punks! But we should not pretend that simply having a more sophisticated model, and a huge data set, absolve us of the statistical problems that have plagued analysts for centuries when testing scientific theories. Our models must be designed precisely to account for possible confounding factors, and we still need controlled studies to carefully assess causality. As computer scientists should know: garbage in, garbage out.
This is not a plea for researchers to 'stay in their lane'. I think criminology and statistics both need fresh ideas, and many of the smartest people I know work in machine-learning. We should all be looking for new areas to apply our ideas in. But working in a new field comes with some responsibility to learn the basic issues in that area. Almost everyone in biology or social science has a story about a physicist who thought they could solve every problem in a new field with a few simple equations, and I don't want data scientists to do the same thing. I fear that if modern data science had been invented before the discovery of the Theory of Gravity, we would now have computers capable of insanely accurate predictions of ballistics and planetary motions, and absolutely no idea how any of it really worked.