Presentations & pubs

Hand-Engineered Image-Computable Models Can Still Outperform DNNs in V1 Similarity

Tanish Mendki, Sudhanshu Srivastava, Ansh Soni

Contributed talk · Vision Sciences Society Annual Meeting · May 2026

Task-based deep neural networks (DNNs) are widely used as models of inferotemporal visual cortex (IT), with early work showing a large jump over previous hand-made models. However, recent work suggests that the relationship between task performance and neural alignment has plateaued and even reversed, with high-performing models becoming worse models of IT. Here, we ask whether this reversal extends to earlier cortical regions. We evaluate a broad set of models, including state-of-the-art IT-similarity models, high-performing CNNs and vision transformers, models designed to reproduce V1, and traditional hand-crafted models such as HMAX. Surprisingly, the most modern models are equal or worse models of V1 than HMAX, even as they are better models of IT. HMAX performs best when similarity measures emphasize representational geometry or strict pairwise matching. These results suggest that hand-crafted image-computable models remain important: in some circumstances, their biologically motivated structure still captures V1 better than modern deep networks.

Hand-Engineered Image-Computable Models Can Still Outperform DNNs in V1 Similarity

Tanish Mendki, Sudhanshu Srivastava, Ansh Soni

Extended abstract · UniReps Workshop, NeurIPS 2025

Humans and Neural Networks Learn to Use Background Structures for Efficient Visual Search

Tanish Mendki, Miguel P. Eckstein

Poster · URCA Poster Colloquium · UC Santa Barbara

Visual search is limited by random variability in the environment, which can make targets difficult to distinguish from distractors. Yet backgrounds can also contain dependable structures that must be learned and efficiently discounted, as in medical images. We investigated how humans and convolutional neural networks use such structure during search. Human observers searched synthetic scenes containing a target among spatially structured Gaussian distractors, while a CNN was trained on the same task and tested after portions of the background were removed. Human accuracy improved across sessions, approaching 75 percent by the end of the second day. The CNN's performance dropped sharply when learned background structure was removed, then recovered once little structure remained and target features alone became sufficient. These findings indicate that both humans and CNNs can learn to treat structured backgrounds as informative cues rather than mere noise, and suggest a way to study how expert observers balance detecting and discounting dependable distractors.

Humans and Neural Networks Learn to Use Background Structures for Efficient Visual Search

Tanish Mendki, Miguel P. Eckstein

Oral presentation · URCA Conference · UC Santa Barbara