Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Downsampled Representations Improve Cross-Model Transfer in hV4 Encoding-Model Metamers

Hayato Ono¹, Ayumu Yamashita², Masataka Sawayama³, Kaoru Amano¹; ¹The University of Tokyo, ²Kobe University, Tokyo Institute of Technology, ³Hokkaido University

Presenter: Hayato Ono

Metamers are physically distinct images that produce the same response in a model or brain area and can therefore probe which image changes leave that response unchanged. In brain-response modeling, encoding models can synthesize images that preserve the predicted response to a reference image. However, matching a single fitted encoding model does not by itself support an inference about shared invariance within an ROI. We therefore define metamer validity as hold-out cross-model transfer to independently fit models not used during synthesis. Using hV4 responses from the Natural Scenes Dataset, we trained six AlexNet-based encoding models, reserving three for synthesis and three for hold-out evaluation. We varied the number of jointly matched synthesis models (K=1,2,3) and compared Pixel and Lowpass synthesis conditions. Hold-out transfer was consistently higher for Lowpass than for Pixel, and increases in K produced more consistent gains in Lowpass than in Pixel. Thus, inferences from encoding-model metamers depend on both multi-model constraints and synthesis parameterization.

Topic Area: Computational Models of Vision & Visual Cortex