Let us classify the state of knowledge into four simple categories: what do we know (known knowns)? What are the limitations of what we know (known unknowns)? What is our degree of certainty about our knowledge (unknown knowns)? What are we unconsciously not exploring (unknown unknowns)? Most research focuses on moving gradually the known unknowns to known knowns with repeated acceptation or refutation of hypotheses.

My question is about the scientific method. Research in my field traditionally relies on hypothesis testing or development, in a deductive or inductive approach. A scientific method relying on hypothesis testing or development is particularly efficient at choosing between possible processes that could explain a phenomena (i.e. to explore the known unknowns), but would fail to scrutinize new areas, the unknown unknowns, whose exploration may lead to totally unexpected scientific discoveries. A way to probe the unknown unknowns (I guess) is to use a data-driven scientific method. The argument is that, with large volume of data describing a situation and data-driven techniques, it becomes possible to discover patterns in the data and the hypotheses that follow. This is a kind of abductive reasoning, where the data-driven algorithm is used to generate hypotheses.

My question is whether a data-driven scientific method can also be useful to explore the known unknowns, as a replacement of the dominant use of hypothesis testing and development? More generally, is it possible to use data-driven algorithms (e.g. machine learning) to produce scientific knowledge, or should they be used only to generate hypotheses from data?

  • 2
    Is this your own classification or is there a source? It is not very clear what "known unknowns" are. Like we know there is a parameter that influences some process, but did not measure a precise value yet? I am not sure how something like that is supposed to "replace" hypothesis testing. One can certainly use data mining and aggregation (including AI techniques like machine learning) not only to detect new patterns but also to parametrize/rank existing models, but this is simply a form of measurement/testing.
    – Conifold
    Mar 11, 2020 at 5:22
  • The knowns and unknowns framework has been popularized by Rumsfeld but probably comes from Avicenna to categorize humans according to their knowledge. I like to use it to classify the state of knowledge because it is very practical but I recognize it is not described in epistemology. The 'known unknowns' are typically what research is about: we know that there is climate change, but the underlying physical mechanisms are mostly unknown.
    – anon
    Mar 11, 2020 at 5:44
  • So, are you asking if a machine programmed to have certain limitations can exceed those limitations?
    – puppetsock
    Mar 11, 2020 at 15:06
  • Welcome to SE Philosophy! Thanks for your contribution. Please take a quick moment to take the tour or find help. You can perform searches here or seek additional clarification at the meta site.
    – J D
    Mar 11, 2020 at 15:52
  • In the context of the philosophy of science, it sounds like you are refering to what Kuhn in SSR calls normal and paradigmatic sciences to an extent. Normal science is full of known unknowns that are determined based on the presumptions of the current paradigm. Finding values of constants, building new tools to measure, collecting data and analyzing to speak to details, where as the concept of the the paradigm shift occurs when someone discovers an unknown and reshapes the entire paradigm.
    – J D
    Mar 11, 2020 at 15:56

1 Answer 1


The notion of science as 'data driven' is misleading. Science is usually either goal driven or anomaly driven (which in the end amount to the same thing).

  • Goal driven science means that someone wants to do something (or maybe just do something better), and they start trying different things as common sense or inspiration dictates. These efforts produce directed data that can then be analyzed, systematized, and used to inform future efforts.
  • Anomaly driven science means that someone notices something unexpected, and sets out trying to figure out what the hell just happened. Again, they start trying different things as common sense or inspiration dictates, which produces directed data that can then be analyzed, systematized, and used to inform future efforts.

known unknowns as you've described them would constitute the goals or anomalies in these two cases: things we believe we ought to be able to do, but currently can't, and things that we observe to be-the-case which we cannot explain. The problem with using machine algorithms to export either case is that machines (currently) lack insight. They cannot distinguish is-expected and is-the-case; they won't recognize anomalies as anomalies. To give a dumb example, say you are digging through a patch of soil and see a momentary flask of purple light. You'd be like: "WTF was that?". You'd recognize it as weird, and as a scientist you'd get curious and start investigating. But how do we program a machine to recognize that something is 'weird', or to be 'curious' about it? We'd have to have our machine take in random information while its going about other tasks and correlate that random information with its own actions, which calls for a fairly high level of self-awareness. I won't suggest it's impossible, but...

  • I understand what you mean by science driven by anomalies and goals, this is well described by Kuhn, Bachelard or Foucault, but the notion of data-driven science is still relevant to me: what if new insights are 'born from the data', that is, a data-driven algorithm interrogates the data to tease subtle correlations that are often inherently invisible to the human because of the multivariate and non-linear nature of the data. The discovered pattern might ultimately serve for the development of new hypotheses.
    – anon
    Mar 12, 2020 at 0:00
  • See also that a similar discussion is provided in Kitchin, R. (2014). Big Data, new epistemologies and paradigm shifts. Big data & society, 1(1), 2053951714528481.
    – anon
    Mar 12, 2020 at 0:08
  • @AlexandreWadoux: My point was that people don't just 'make data.' Data only occurs from within structure and context. Data that occurs outside of some structure and context is what we commonly refer to as 'noise.' There's an analogy to music, here; you don't have music without composition. Mar 12, 2020 at 1:01
  • It is true that producing data never occur in a 'scientific vacuum', but in many fields data are now recorded without any goal. This is the case, for example, for genome sequencing or remote sensing images. In this sense, the data are not used to test a hypothesis or evaluate a theory, but are the driving forces behind the analysis.
    – anon
    Mar 12, 2020 at 1:15

You must log in to answer this question.