Amount awarded: $7,400

Abstract: Neuroscientific investigations of neural representations often proceed by identifying patterns of neural activity that correlate with an experimental parameter and, ideally, cause experimentally relevant behavior. Such desiderata as these provide evidence that the pattern carries information or is content bearing about the relevant cognitive or behavioral explanandum. Because neural representations are theoretically discrete units, a challenge neuroscientists face is in individuating those content-bearing patterns from other structures and processes.

When researchers present results that purportedly provide evidence for neural representations, the results show the presence of some property that is contrasted with conditions in which that property is absent. For example, researchers might claim that a pattern of activity in the temporal lobe represents a consciously perceived face because such a pattern is strongly activated while a face is seen, but not when it is unseen despite constant sensory stimulation (Tong et al., 1998). The natural response from opposing interlocutors is that the pattern in question is not necessarily absent when the face is unseen: with a different tool or theory of the brain, the relevant property could still have been detected throughout all the conditions (Hebart & Baker, 2018).

One way to make progress on the question of which properties are relevant for serving as a neural representation is to develop tools with which to determine whether a neural representation is not present. Developing a novel means to provide evidence for the absence of a representation would push many debates forward as well as improve theorizing about neural representations. We propose to leverage machine learning to provide novel means for detecting the properties of interest and determining when they are absent.

Training machine learning algorithms on neural data has allowed researchers to decode activity and predict some cognitive or behavioral output from neural activity. Our project uses classifiers trained on neural data to provide an estimate of the absence of a particular pattern in that data. In particular, we train numerous classifiers systematically varying processing pipelines to create classifiers with varying sensitivity to the pattern. Mapping out this classifier-space will provide an estimate of the sensitivity at which a pattern or its absence is detected. This tool feasibly provides evidence for an interpretation of data suggesting that a pattern exists only at a particular time or place and not elsewhere in the data.

Description: Neuroscientists are interested not only in finding meaningful signals in the brain through analysis of data, but also in showing that those signals are absent in other parts of the data. Such absence can support interpretations that the signal does not occur in another place or at another time. We develop a tool to measure the strength of evidence that a signal is recoverable from the data by leveraging degrees of classifier sensitivity to the signal.