Over history, major discoveries have often been marked by new, frequently ingenious, methods of representing knowledge—the periodic table, the tree of life, Feynman diagrams, dividing time into geological or historical periods, abstract vs representational art, weather maps, language families. These representations, whether subtly or obviously, embody certain assumptions. In some cases, they are way stations on the road to a better understanding—the periodic table hid the germ of the concept that became atomic orbitals, the Linnaean classification hid the germ of descent with modification. Compelling representations can be a barrier to progress—notoriously the representation of planetary orbits by nested epicycles. Others further progress—representing planetary orbits by ellipses spurred Newton’s spectacular advance in physics and mathematics.
We are at a moment when it has suddenly become possible to investigate the issue of knowledge representation more profoundly.
An interdisciplinary team of distinguished researchers from psychology, neuroscience, biology, computer science, mathematics, sociology, the humanities and history of science have come together to investigate knowledge representation on multiple levels and from a diversity of viewpoints:
- How does the choice of available methods of representation influence what is investigated and the ability to make major advances, both in contemporary research and historically?
- How can one best quantify efficiency of different representations of knowledge?
- How does choice of representation affect the ability to retrieve a given piece of knowledge in the brain, in the cell, for researchers in a research community, for those outside a research community who are unfamiliar with the representation?
- Are there new families of methods for representing knowledge that would be especially well-suited to fields where large complex datasets are available? How does entrenchment of existing knowledge representations impede the introduction of radically different methods?
- What does the “space” of all knowledge representations of a given collection of data look like? Can one optimize over this space for a chosen criterion—compactness of representation, ease of retrieval, accessibility to those outside the field, learnability? Is there a lower-dimensional subspace of representations on which the most useful representations usually are found?
- What is the relationship between method of knowledge representation being used and possible hypotheses for experiments? Does this produce in some cases a natural prior of the space of possible hypotheses?