Research · Tanish Mendki

Current research · Columbia University

How do we measure representational alignment between models and brains?

Comparing representations across brains and models has become central to cognitive computational neuroscience. Many increasingly diverse sets of measures exist for this purpose. Each measure is invariant to a different set of transformations, and so treats different properties as signal or nuisance variation. In order to successfully apply them in NeuroAI research, I'm interested in understanding the underlying assumptions and limitations of these measures.

Ongoing collaboration · UC Santa Barbara, University of Pennsylvania, UC San Diego

What makes a vision model brain-like?

A vast variety of architectures and models have been shown to exhibit a similar degree of alignment with biological vision. When multiple models are indistinguishable in terms of alignment, how do we determine which one is more brain-like? Our results show that carefully chosen inductive structure can outperform architectural scale in V1 similarity. I'm especially interested in identifying the model properties that drive alignment, rather than treating a single benchmark score as an explanation.

Past research · UC Santa Barbara

How do humans and neural networks perform visual search?

Visual search rarely happens against a blank background. This project asked whether networks learn to use regularities in structured scenes as humans do—or whether they solve the task through shortcuts that bypass the target itself. I built a synthetic-image pipeline and model-introspection framework, trained multiple CNN architectures to high accuracy, and tested what information actually supported their decisions. The models relied strongly on background context, revealing both an efficient strategy and an important point of divergence from human perception.

Past research · UC Santa Barbara

Can humans and AI recognize deceptive visual explanations?

Fluent explanations can sound authoritative even when they are wrong. We curated 360 images with deliberately misleading descriptions across five knowledge domains, collected human judgments, and compared them with evaluations from GPT-4 and Gemini 2.5. We were interested in the gap between visual understanding and plausible language: when an evaluator notices that a description conflicts with an image, and when surface-level expertise is enough to persuade it.