Group items tagged

Roughly put, building a machine-learning model involves training it on a large number of examples and then testing it on a bunch of similar examples that it has not yet seen. When the model passes the test, you're done. What the Google researchers point out is that this bar is too low. The training process can produce many different models that all pass the test but-and this is the crucial part-these models will differ in small, arbitrary ways, depending on things like the random values given to the nodes in a neural network before training starts, the way training data is selected or represented, the number of training runs, and so on. These small, often random, differences are typically overlooked if they don't affect how a model does on the test. But it turns out they can lead to huge variation in performance in the real world. In other words, the process used to build most machine-learning models today cannot tell which models will work in the real world and which ones won't.

Interesting discussion of why headlines like this one "85% accurate" for the detection of stomach cancer" about an experimental breath test are problematic (because some people who don't have the condition get diagnosed with it, and they can miss people who genuinely do have the condition!). Good example using pregnancy tests as an infographic.

Maths quiz. If you take a Covid test that only gives a false positive one time in every 1,000, what's the chance that you've actually got Covid? Surely it's 99.9%, right? No! The correct answer is: you have no idea. You don't have enough information to make the judgment.

The proposal to drug test welfare recipients keeps on bouncing back. The most recent attempt, announced last week, is now the third proposal since 2017. But the tenacity with which the government is pursuing this agenda reflects, not necessarily a fixed policy position, but rather a moral stance. And this moral stance conflicts with that of the proposals' critics. Are we doomed to countless repeats of the same policy proposal? Or, as the Australian Social Policy Conference heard in Sydney this week, can we use philosophical arguments to help break the deadlock?

During the 2016 EU referendum campaign, both sides used statistics pretty freely to back their arguments. Understandably, UK broadcasters felt compelled to balance competing perspectives, giving audiences the opportunity to hear the relative merits of leaving or remaining in the EU. In doing so, however, the truth of these statistical claims was not always properly tested. This might help explain some of the public's misconceptions about EU membership. So, for example, although independent sources repeatedly challenged the Leave campaign's claim that the UK government spent £350m per week on EU membership, an IPSOS MORI survey found that almost half of respondents believed this was true just days before the election. Of the 6,916 news items examined in our research, more than 20% featured a statistic. Most of these statistical references were fairly vague, with little or limited context or explanation. Overall, only a third provided some context or made use of comparative data. Statistics featured mostly in stories about business, the economy, politics and health. So, for example, three-quarters of all economics items featured at least one statistic, compared to almost half of news about business. But there were some areas - where statistics might play a useful role in communicating trends or levels of risk - that statistics were rarely used.

But shouldn't there be some kind of accountability when you're talking about public money? My response would be what kind of accountability do we have now? Our schools are graded by test scores, and half the schools are doing worse than the other half and nobody does anything about that at all. This is the advantage of letting parents control the venue for education. And Michael Gove (excerpt at) https://twitter.com/Annelies_Leeuw/status/851945727110643717/photo/1

As students from privileged backgrounds flock to schools that were supposed to be the ultimate symbol of egalitarianism, experts fear they may be reinforcing class and cultural divisions

A new paper published in PLOS One by John Cook, Stephan Lewandowsky, and Ullrich Ecker tests the power of inoculation; not against disease, but against the sort of misinformation that created the conditions leading to Minnesota measles outbreak. Inoculation theory suggests that exposing people to the tricks used to spread misinformation can equip them with the tools to recognize and reject such bogus claims.

Numbers are often used to persuade rather than inform, statistical literacy needs to be improved, and so surely we need more statistics courses in schools and universities? Well, yes, but this should not mean more of the same. After years of researching and teaching statistical methods, I am not alone in concluding that the way in which we teach statistics can be counterproductive, with an overemphasis on mathematical foundations through probability theory, long lists of tests and formulae to apply, and toy problems involving, say, calculating the standard deviation of the weights of cod. The American Statistical Association's Guidelines for Assessment and Instruction in Statistics Education (2016) strongly recommended changing the pedagogy of statistics into one based on problemsolving, real-world examples, and with an emphasis on communication.

Interesting discussion of evidence on what does, and does not, work in tackling gender bias in recruitment and management processes. Evidence shows that skills-based assessment tasks (where candidates are given tests that replicate the work they'll actually do on the job) and structured interviews (where all candidates are given the same questions in the same order) have a positive impact on diverse recruitment. Unstructured interviews are more likely to allow unfair bias to creep in. Making promotion and pay processes more transparent can reduce pay inequality: when decisions are reviewed by others, managers realise they need to be objective and evidence-based. Evidence also shows women ask for less money than men. To encourage them to negotiate more, employers should make the possible salary range for roles clear. Studies indicate that women are put off negotiating when they're not sure what a reasonable offer is. "A lot of employers are genuinely really keen to reduce the gender pay gap, and also want to show they're making a change. But they're starved for information about what is likely to work,"

As Omicron surges through the community, getting the right answer from a Rapid Antigen Test (RAT) is not as straightforward as reading one or two lines off the kit. RATs are a convenient diagnostic tool to detect COVID virus fragments in nasal secretions or saliva. They are designed to be self-administered and give an answer in minutes. Detecting infection early is critical to preventing spread and allowing persons at risk of severe disease to get timely access to close monitoring and new life-saving therapies. As governments plan to distribute tens of millions of RAT kits to schools and workplaces in coming weeks to help Australians work and study safely, it is important that we understand how to best use this diagnostic tool to reduce transmission and unnecessary disruptions to our lives and economy.

Almost one in four cancers detected in men were overdiagnosed in 2012, according to our new research, published today in the Medical Journal of Australia. In the same year, we found that approximately one in five cancers in women were overdiagnosed. Overdiagnosis is when a person is diagnosed with a "harmless" cancer that either never grows or grows very slowly. These cancers are sometimes called low or ultra-low-risk cancers and wouldn't have spread or caused any problems even if left untreated. Cancer overdiagnosis can result in people having unnecessary treatments, such as surgery, radiotherapy and hormone therapy. Being diagnosed with cancer and having cancer treatments can cause physical, psychological and financial harms.