Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Selective Is Selective, or Not? Investigating Consistency and Task Relevance of Selectivity Metrics in DNNs

Anastasia Lado¹, Katharina Dobs¹; ¹Justus Liebig Universität Gießen

Presenter: Anastasia Lado

Neural response selectivity is a long-standing phenomenon in cognitive neuroscience, with distinct cortical areas selectively responding to visual categories across processing stages. Such selectivity is typically quantified using measures ranging from simple response ratios to detailed statistical comparisons between preferred and non-preferred categories. But how consistent and stable are these measures? And, critically, does selectivity capture behavioral relevance? Recent computational studies have shown that both trained and untrained deep neural networks (DNNs) exhibit category-selective units. Here, using face selectivity as our test case, we leverage DNNs to systematically compare a broad range of selectivity metrics while assessing their relevance to task performance. Our results reveal low agreement between selectivity metrics and lesioning-based rankings, and the consistency among metrics varies with spatial scale, processing stage, and training. However, all metrics yield similar face decoding accuracy. These findings caution against overreliance on any single metric and inform the interpretation of selectivity in computational and neural data.

Topic Area: Object Recognition & Visual Attention

Extended Abstract: Full Text PDF