WHAT YOU SHOULD KNOW
The big data fallacy has plunged scientific research into crisis.
If you’re a little familiar with the world of computation, you’ll have heard of Moore’s law: the idea that the raw computing power of our devices doubles every two years. Since its first formulation in 1965, this law has held approximately true. But as you can probably attest from casual observation, the fact that technology is getting ever smaller and faster doesn’t necessarily mean that it’s making our lives any easier.
Tech optimists who cite Moore’s law usually do so in the spirit of computational optimism. They believe that more computation is always better, and that the more data we collect and process of the world, the more our understanding of it increases.
In science, this belief has led to a research system that values automated testing that generates tons of data over more messy human empiricism. In drug research, for example, the role of human scientists is now often reduced to programming and overseeing the work of machines engaged in a process called High-Throughput Screening. The computer tests the effects and interactions of thousands of chemical compounds a day, in the hope that it will eventually stumble upon a combination that is useful in treating a particular disease.
The problem is, this approach doesn’t seem to work at all. Every nine years since the 60s, the number of new drugs approved for human use per billion US dollars in spending on research and development has halved. Commentators have cynically started calling this effect Eroom’s law — Moore’s law backwards.
The big data fallacy that encourages quantity over quality is palpable in all of science. While the number of scientific studies, journals, and papers has been steadily increasing over the past decades, so has the number of mistakes, plagiarism and fraud in scientific research. Experts are increasingly talking about the replication crisis of modern science. This refers to the fact that when many scientific studies are conducted a second time by a different group of researchers, they cannot reproduce the original results.
In 2011, for instance, the University of Virginia reran five landmark cancer studies of recent years. Only two experiments could be successfully replicated; two others were inconclusive; and one failed completely.
Even as scientific research is gathering more and more data about the world, the pace of scientific discovery is actually slowing. Instead of endowing us with a better grasp on the world, the current overflow of information is negatively affecting our ability to process what’s going on around us.
As a tool of capitalism, technology drives inequality.
On the surface, Slough is a pretty unremarkable, small town some 25 miles outside of London. Unbeknownst to most, however, the many vast and anonymous warehouses that line its main road form the physical base of some of the most important parts of our digital world. For example, one of them, given the unassuming name LD4, houses the data servers of the London Stock Exchange.
The fibre optic cables that lead to and from LD4 carry financial information of almost unimaginable value, transporting and receiving it to and from other financial data centers around the world at the speed of light. This super-speedy network between companies, investors and markets has given rise to a new type of financial exchange: high-frequency trading.
Today, financial traders can react almost instantly to drops and spikes in the market. To do so in a matter of milliseconds, they enlist the help of algorithms and bots that monitor prices, make mock offers and shadow transactions to confuse other traders, and even scan and interpret news headlines to anticipate the economic effects of major events around the world.
But it’s becoming more and more apparent that even insiders can’t keep up with the logic of their computers in the hyper-accelerated world of digital finance. For instance, on May 10, 2010, the Dow Jones experienced an unprecedented 600 point crash — an equivalent of six billion dollars lost — then suddenly recovered minutes later. Such flash crashes are growing more common, and no human is able to pinpoint what exactly causes them.
While machines increasingly confuse humans in some areas, they replace us outright in others. Just take behemoth Amazon, which is already using fleets of robots to store, sort, and pick out products. Where it still uses human ‘pickers’, Amazon does so out of financial incentive and essentially treats those workers like robots: The workers are guided and monitored via a handheld device, which sends them to different locations in the warehouse in a manner that maxes out efficiency and minimizes socializing with coworkers.
Worryingly, politicians and companies are providing little perspective on what social security system could replace full-time employment. And so technology, far from being the great equalizer that we’ve been promised, is just another tool for concentrating power in the hands of the few.
Machine learning encodes the bias of our past and carries it into our future.
There’s a lore about an AI built by the US Army that illustrates the dangers and limits of machine learning, by which is meant: teaching computers how to think.
Allegedly, the army tried to train a computer to recognize camouflaged tanks in a forest. To do so, they presented it with picture after picture of forests with tanks hidden in them, and picture after picture of forests with no tank, until the AI had learned to tell the test images apart perfectly. But out in the field, the AI failed completely. It was no better than a human at guessing whether there was a tank in a particular forest or not.
Someone noticed only later, that all the training photos with tanks had been taken on a sunny day, and all those without a tank had been taken on a cloudy day. The machine hadn’t learned the difference between a forest with a tank, and a forest without one: it had learned the difference between a sunny and a cloudy day.
This story shows us that when we train machines to think, we cannot expect them to think like us. In many cases, we might never be able to understand how or why they reached their conclusion at all. Computers build their own multidimensional simulations of the world that are entirely different from our human experience of it.
But more than just an idea worth considering, the mysterious nature of a machine’s mind can also be used to justify the conclusions they come to, even when these are controversial or dangerous.
In 2016, two researchers from a University in Shanghai made a stir when they claimed to have developed software that could tell the difference between a criminal and a non-criminal face. When their experiment was criticized on the basis that the software would surely over-represent marginalized communities, they claimed that they had constructed it purely for academic purposes, and that it, and machine learning in general, was inherently “free of bias.”
The idea that algorithms and computation are unbiased is shared by many AI enthusiasts. What they fail to acknowledge is that machines tend to be trained with data, and the only data we have is of our past. Since our past is rife with violence, injustice, and racism, whether we intend to or not, the machines we train with this data are going to replicate that violence, injustice, and racism and project them into the future.
As recently as a few years ago, for example, Asian-Americans tried in vain to take family photos with their Nikon Coolpix S630. Instead of taking a picture, the “smart” camera repeatedly displayed the error message “Did someone blink?”
Olusola Bodunrin is a social commentator with a passion for gathering facts on social issues. He is writer, businessman, blogger, and wardrobe consultant. He is a graduate of Philosophy from the University of Ado-Ekiti. He anchors – ‘What You Should Know’ on SHEFFA.
‘What You Should Know’ is a column that offers to educate and enlighten the public on general falsehood and myths.