This commentary relates to the Frontiers in Neuroscience article “The Value of Failure in Science: The Story of Grandmother Cells in Neuroscience” by Ann-Sophie Barwich. This commentary is not against the idea of failure analysis, but about the argument that grandmother cells is a failed notion. I summarize first the discussion in this article by Barwich about grandmother cells (and related concepts—concept cells, associative learning, hierarchical processing, sparse coding, and so on) in the sequence they were presented (all references are in the Barwich article): (1) that they respond to a specific but complex stimulus; (2) that they account for selectivity and specificity of neural responses and localization and convergence in neural hierarchical processing; (3) that localization as a paradigm continues to persist; (4) that according to Marr (1982), answers about mind and brain could not be found at the single cell level; (5) that the framework of convergent, hierarchical, and localized signal integration has not been successful; (6) however, even after this general characterization, it's still unclear to Barwich what grandmother cells actually were; specifically, to what processes they referred to; (7) that grandmother cells were opaque both on a theoretical and a practical level; (8) that it is undetermined what kind of information was processed and integrated with such a hypothetical neuron; do they only respond to visual input or also process cross-modal cues including auditory and olfactory signals; if they do cross-modal integration, where is the centrum of such integration; (9) that the broader question is how does unified phenomenal experience arise from separated and specialized neural processes; (10) that the scope of the grandmother cell concept was not clear-cut; do they respond only to a particular individual (a grandmother) or to a category (grandmothers); (11) she then introduces the notion of concept cells through Konorski's gnostic units, which were associated with recordings of localized responses to complex objects, such as faces; (12) argues that although concept and grandmother cells and gnostic units seemed sufficiently similar, they are not because they have different processes; (13) then she illustrates concept cells (Quiroga et al., 2008)—single neurons that respond selectively to various visual representations of individual objects (Sydney Opera House) and people (Jennifer Aniston); (14) then introduces sparse coding (activation of small groups of neurons) as an explanation for specialized neural responses (e.g., faces); (15) argues that sparse coding is not exclusive to a specific stimuli, such as Halle Berry; (16) that sparse coding is a different theory of neural representation—sets of neurons build a net of neural representation, potentially unrelated in their feature coding pathways (e.g., visual input integrated with auditory input); that these sets of neurons do not code concepts strictly bottom-up from simple signals to complex individuals and categories; (17) that the main difference between grandmother cells and sparse coding is that grandmother cells emerged from the hierarchical coding hypothesis, but sparse coding doesn't require hierarchical coding; (18) that sparse coding uses a separate theory of learning, that of associative learning; (19) and claims that it's debatable whether abstraction defines the ultimate nature of the brain.

Commentary - There are many flawed arguments in the analysis presented by Barwich as summarized above. Here are some of the main ones.

Overall, Barwich's analysis lacks rigorous comparison of ideas, uses flawed arguments and has false claims. In addition, characterizations of concept cells by Reddy and Thorpe (2014) and others confirm the prediction of Barlow (2009) that grandmother cells exist and “can now be recorded from and studied reliably.” One should also note that grandmother cells are just a special type of complex abstract cell. Multisensory neurons provide extensive evidence for other types of abstract cells in the brain (Roy, 2017).