By:
Philipp Haueis, Department of Philosophy, Bielefeld University, Bielefeld, Germany
Daniel S. Margulies, Oxford Centre for Integrative Neuroimaging (OxCIN), FMRIB, Nuffield Department of Clinical Neurosciences, University of Oxford
You can find Philipp’s and Daniel’s paper (open access) at the European Journal of Neuroscience here: Developing Concepts for Neuroscience: A Philosophical Toolkit
Abstract
Alongside models and methods, concepts are crucial tools to study and understand the brain. They help us pursue various goals, such as describing phenomena based on patterns in the data or explaining why these phenomena occur. Yet while terms such as “action potential” or “network” guide our efforts to reach these goals, other concepts have failed to advance our understanding of the brain. In this paper, we draw on recent work from philosophy of science to show that the success or failure of concepts in neuroscience depends on the epistemic goals the field aims to achieve. Looking at cases such as “default mode network,” “cortical column,” and “hierarchy,” we formulate conditions under which introducing, refining, or replacing a concept succeeds or fails. These cases suggest that to better evaluate our concepts, we should make explicit which goals we aim to achieve when using them.
Commentary from Philipp and Daniel
Alongside methods and models, concepts are crucial tools for neuroscience. We use terms like “neuron”, “action potential” or “network” to identify which parts and processes in the brain are functionally important and thus worthy of investigation.
While these concepts advanced neuroscientific understanding, not every conceptual development succeeds. An example is “grandmother cell”: it failed to explain how the cortex processes sensory information and gave way to alternatives like sparse coding. But what distinguishes success from failure?
In our recent paper, we argue that we cannot talk about success or failure of concepts in neuroscience without taking the community’s epistemic goals into account. We introduce, revise or replace concepts to better describe neural phenomena, classify them into kinds, and explain why or predict when they occur. Successful concepts improve our understanding of the brain because they help us achieve these goals; failed concepts do not.
Our paper analyses cases such as “default mode network”, “hierarchy” and “cortical column” to formulate conditions under which introducing, refining or replacing concepts succeeds. “Default mode network”, for instance, reshaped how neuroimaging studies cognitive function. It denotes a set of areas that unexpectedly decreased their activity during goal-directed tasks. When does introducing a novel concept adequately characterize such an unexpected pattern?
We propose two conditions: a newly introduced concept should be anchored in a property that is significant for pursuing an epistemic goal, and open-endedly applicable in future research. “Default mode network” meets only one of these. Because “default mode” was defined by a metabolic baseline, the concept was not anchored in a property significant for describing the system’s cognitive function. But it was open-ended, allowing researchers to explore the network’s roles, e.g. in episodic memory, self-related processing or representational abstraction.
Concepts are also often refined over the course of inquiry. Classic studies using “hierarchy”, for instance, classify cortical areas by the information they represent (e.g. V4 represents color). Newer studies of network topology, however, use “hierarchy” to describe an area’s influence on the network, without invoking representational content. How should researchers relate these meanings?
The answer depends on epistemic goals, and each option changes our understanding in a different way. If researchers emphasize converging data, they may view graph-based methods as adding substance to representational hierarchy, leaving this understanding relatively intact. If researchers weigh conflicting data more heavily (e.g. V4 is mid-level representationally, but highly central topologically), they may replace the representational understanding with the topological one. Both options assume that old and new studies pursue the same epistemic goals. But they may also pursue different goals, e.g. explaining actual signal processing vs. efficient communication under wiring constraints. In that case, pluralism about “hierarchy” would be the better option.
Finally, concepts become outdated when they no longer serve the central epistemic goal for which they were introduced. “Cortical column” was introduced by Vernon Mountcastle in the 1950s to describe a basic unit of the mammalian neocortex. It guided research that won Hubel and Wiesel the Nobel Prize in 1981, but newer studies suggest that it refers to diverse kinds of structure rather than a basic building block. If we keep searching for a building block, we should replace “cortical column” with alternatives like “canonical microcircuit”. If we abandon this goal, we should retire “cortical column”, acknowledge its past role, while also searching for new concepts.
Our framework suggests that to better evaluate when concepts succeed or fail, we should make explicit which epistemic goals we aim to achieve when using concepts in neuroscientific practice.
Leave a Reply