Unlike what happens when we want to know how long it takes a stone to reach the ground, there are no safe answers in the realm of human life. If there were, it probably wouldn't be about life, but about a stone. In biomedicine, it is difficult to know if an effect is real since the standards differ between the different fields. Not all tools work with every problem, and the levels of complexity of our knowledge vary even before starting a study.
Despite this, a common element in the different areas of biomedicine is the possibility of replicating, through new studies, what has already been observed in a first investigation. For years we have been discouraged to do so: why waste money repeating what someone has already done before? Now, however, numerous researchers are realizing that it is not feasible to exclude replication studies.
For these to work, it is essential to have a detailed explanation of how the original work was done. This includes instructions, raw data and even some of the software designed for the occasion. In the past, scientists have been reluctant to share this information. But the situation is changing: science is a common enterprise and we must be open and collaborative.
Like any other science, linguistics is based on the scientific method. One of the main objectives of this discipline is to analyze the languages of the world to discover what is possible and what is not in human language. From there, experts try to progress towards the goal of understanding human cognition through our ability to speak. The question of objectivity and scientific "truth" is related to the investigation of threatened languages. In a sense, the truth depends on the context: what we consider true today may change as we get more data or clues that improve our methods. The investigation of endangered languages often reveals things we did not believe possible in human language. That forces us to reexamine previous statements about their limits, so that, on occasion, what we thought was true may cease to be.
Linguists use a series of criteria to identify those languages in danger of disappearing and determine to what extent they are: Are there still children who learn it? How many people speak it? Is the percentage of speakers decreasing with respect to the general population? Are the contexts in which it is used being reduced? Thus, there is a certain urgency to describe the endangered languages and document them while they are still used, as they help us determine the range of possibilities from the linguistic point of view. Currently, about 6500 languages are known, and about 45 percent of them are at risk of becoming extinct.
The basic unit of truth in paleobiology is the fossil, an unequivocal testimony of the past life, but we also use genetic tests obtained from living organisms to locate the fossils in the tree of life. Together, these data help us understand the evolution of life forms and the links that they maintain with each other. Since we study extinct animals that lived immersed in a larger ecosystem, we take information from other fields: the chemical analysis of the surrounding rocks allows us to get an idea of the fossil's age, where the continents could have been at that time, what environmental changes were happening, and so on.
To discover the fossils, we comb the landscape to glimpse them among the rocks. It is possible to distinguish one from an ancient stone by its morphology and its internal structure. A fossilized bone contains tiny cylinders, called osteones, where blood vessels pass through the bone tissue. Some fossils are unmistakable: the entire femur of a dinosaur is a bone of gigantic dimensions. Small fragments can also be revealing. In mammals, my subject of study, it is possible to find out a lot of information from the morphology of a single tooth. And we can combine that data with the genetic ones. We obtain these from DNA samples of living species that we believe are related to fossils, by virtue of anatomy and other evidence.
We not only undertake these investigations to reconstruct the world of the past, but also to find out what they can tell us about the contemporary world. 55 million years ago there was a huge rebound in temperature, nothing comparable to the current one, but still, we have discovered drastic changes in the fauna and flora of the time. The comparison of these changes could tell us how the current bodies related to those of that period could respond to the current climate change.
The most important epistemological issue in the field of machine learning is: what capacity do we have to prove a hypothesis?
Algorithms learn to detect patterns and details from huge sets of examples; They identify a cat like this after seeing thousands of photographs of them. As long as we do not have a greater capacity for interpretation, we can prove how a result has been reached using the conclusions of the algorithms. This makes the ghost appear that we have no real responsibility for the results of deep learning systems, not to mention their effects on social institutions. Such subjects are part of a lively debate in this field.
Another question that arises is: does machine learning represent a kind of rejection of the scientific method, which advocates finding, not only correlations but also causalities? In many machine learning studies, the correlation has become the new guiding principle, leaving aside causality. That raises real questions about verifiability.
In some cases, we may have to step back, as in the field of automatic vision and recognition of emotions. These are systems that extrapolate from photographs of people to predict their race, gender, sexuality or possibility of being criminal. These approaches worry from a scientific and ethical point of view (due to the influence of phrenology and physiognomy). Focusing on the correlation should raise deep suspicions about our ability to make claims about people's identity. That is a very strong statement, but given the decades of research on these issues in the humanities and social sciences, it should not be controversial.
In statistics, we do not see the entire universe, but only a portion of it.
A small portion, usually, that could tell us a completely different story from that of another small portion. We try to move from those fragments to a more general truth. Many consider that that basic unit of truth is the p-value, a statistical measure of how surprising it is that we observe in our small portion, assuming that our premises about the universe are valid. But I don't think that is correct.
In reality, the notion of statistical significance is based on an arbitrary threshold for the p-value and may have little relation to substantive or scientific significance. It is too easy to fall into a mental scheme that assigns meaning to that arbitrary threshold: it gives us a false sense of certainty. And it is also too easy to hide a multitude of scientific sins behind the p-value.
One way to strengthen the p-value would be through a cultural shift towards transparency. If, in addition to communicating the p-value, we also show how we have reached it (the standard error, the standard deviation or other measures of uncertainty, for example) we can better reflect what that figure means. The more information we publish, the harder it will be to hide behind the p-value. I don't know if we will succeed, but I think we should try.
The public assumes that the mere existence of data implies that these data are true. But the reality is that all data is contaminated. It is the people who generate the data and, therefore, these present defects, just like the people. One of the roles of data journalists is to question the premise of truthfulness, which is an important task of responsibility - a control tool to make sure that we are not being dragged through the data collection and that we are not taking inappropriate social decisions.
To question the data we must carry out a huge cleaning job. You have to debug them and organize them: you have to check the math. And we must also admit uncertainty. If you are a scientist and do not have data, you cannot write your article. But one of the wonderful aspects of being a data journalist is that the lack of data does not deter us - sometimes I draw equally interesting conclusions from the lack of data.
The control one has in experimental sciences is much stronger than in behavioral science — the ability to detect small effects on people is much less than, say, chemistry. Not only that; Human behavior changes over time and according to culture. When we think about the truth in behavioral science, it is not only very important to reproduce a study directly but to extend its reproducibility to a large number of situations: field studies, correlational, longitudinal studies, and so on.
So how do we measure racism? Something that is not an individual behavior but a pattern of consequences, an integral system that oppresses people. The best approach is to observe the behavior patterns and then see what happens when a variable is altered or controlled. How does the pattern change? Let's take order maintenance. If we remove the prejudice from the equation, the absurd racial pattern persists. The same goes for poverty, education and a multitude of things that we believe predict crime. None is enough to explain the patterns of police actions conditioned by the racial issue. That means we still have work to do. Because it is as if we do not know how to maintain order without exercising unnecessary violence and with an equal attitude. Let's just look at the suburbs.
Of course, there is uncertainty. In almost nowhere in this world are we close to relying on causation? Our responsibility as scientists is to limit these uncertainties since an erroneous calculation in what leads to something like racism means the difference between exercising a correct police function or one that is not.
Science does not seek the truth, as many believe. Its true purpose is to ask better questions. We conduct experiments because we ignore something and want to know more, and sometimes those experiments fail. But what we learn from ignorance and failure raises new questions and new uncertainties. And they are better questions and better uncertainties that give rise to new experiments. And so on.
Let's look at my field, neurobiology. For about fifty years the fundamental question about the sensory system has been: what information is sent to the brain? For example, what do our eyes tell the brain? Now we study this idea but vice versa: the brain actually asks the sensory system questions. In this way, the brain may not only be producing a huge amount of visual information from the eye; on the contrary, it asks the eye to look for specific information.
In science, there are always loose ends and small dead ends. When we might believe that everything is resolved, something new and unexpected always comes up. But uncertainty is valuable. I should not create anxiety. It is an opportunity.
Physics is the most mature science and physicists are obsessed with the truth. Out there extends a real universe. The main miracle is that there are simple underlying laws, expressed in the precise language of mathematics, that allow us to describe it. That said, physicists do not work with certainties, but with degrees of confidence. We have learned the lesson: throughout history, we have discovered time and again that some principle that we considered key to the definitive description of reality was not entirely correct.
To clarify how the world works, we formulate hypotheses and theories and build experiments to corroborate them. Historically, the method works. For example, physicists predicted the existence of the Higgs boson in 1964, built the Large Hadron Collider (LHC) between the late nineties and early this century and found the particle in 2012. Other times the experiment cannot take out: it is either too large or expensive, or it is not viable with the available technology In such cases, we propose mental experiments based on the infrastructure of existing mathematical laws and experimental data.
Here is an example. The concept of spacetime has been accepted since the beginning of the 20th century. However, to observe it on a small scale, we need a great resolution. This is the reason why the LHC has a ring of 27 kilometers in length: to generate the enormous amount of energy it takes to probe the tiny distances between particles. However, there comes a time when this strategy fails. If we accumulate too much energy in a space small enough, we will end up creating a black hole. Our attempt to see what happens at these distances makes that goal impossible, and the notion of spacetime crumbles.
At any moment in history we have managed to understand some aspects of the world, but not all. When a revolutionary change brings more elements to the global image, we are forced to restructure what we already knew. The old is still a piece of truth, but we have to turn it around and fit it in another way in a larger puzzle.
