Doby Rahnev
@dobyrahnev.bsky.social
Perceptual decision making, visual metacognition, computational cognition, cognitive neuroscience, neuroAI. Associate professor at Georgia Tech. Director of Computations of Subjective Perception lab: https://rahnevlab.gatech.edu
This paper started almost a decade ago in collaboration with the amazing @racheldenison.bsky.social. Marshall Green and Mingjia Hu did the actual work.
February 5, 2026 at 3:01 PM
This paper started almost a decade ago in collaboration with the amazing @racheldenison.bsky.social. Marshall Green and Mingjia Hu did the actual work.
These results don't mean that ANNs are a good model of internal evidence for all visual tasks (far from it), but they do show that this is likely to be the case for simple visual spaces.
February 5, 2026 at 3:01 PM
These results don't mean that ANNs are a good model of internal evidence for all visual tasks (far from it), but they do show that this is likely to be the case for simple visual spaces.
Critically, artificial neural networks (ANNs) trained on the orientation task reproduced both the fine- and coarse scale results as emergent properties, without any special training or fine-tuning. This was the same for 3-, 4-, and 5-layer networks.
February 5, 2026 at 3:01 PM
Critically, artificial neural networks (ANNs) trained on the orientation task reproduced both the fine- and coarse scale results as emergent properties, without any special training or fine-tuning. This was the same for 3-, 4-, and 5-layer networks.
At the same time, increasing the stimulus tilt in coarse-scale increments had a highly non-linear transformation with a plateau beyond 14 degrees. This difference between fine- and coarse-scale results isn't predicted a priori from most standard models.
February 5, 2026 at 3:01 PM
At the same time, increasing the stimulus tilt in coarse-scale increments had a highly non-linear transformation with a plateau beyond 14 degrees. This difference between fine- and coarse-scale results isn't predicted a priori from most standard models.
In a task where subjects judged if Gabors were tilted clockwise or counterclockwise, we examined how orientation is transformed into internal evidence. We found that increasing the stimulus tilt in fine-scale increments resulted in a linear increase in sensitivity.
February 5, 2026 at 3:01 PM
In a task where subjects judged if Gabors were tilted clockwise or counterclockwise, we examined how orientation is transformed into internal evidence. We found that increasing the stimulus tilt in fine-scale increments resulted in a linear increase in sensitivity.
Great work by the whole team: Medha Shekhar, @herrickfung.bsky.social, Krish Saxena, and Farshad Rafiei. Code and data posted as always.
January 26, 2026 at 7:18 PM
Great work by the whole team: Medha Shekhar, @herrickfung.bsky.social, Krish Saxena, and Farshad Rafiei. Code and data posted as always.
More generally, our work represents the power of ANNs to uncover how humans represent and operate on perceptual information.
January 26, 2026 at 7:18 PM
More generally, our work represents the power of ANNs to uncover how humans represent and operate on perceptual information.
We found clear evidence that the Top2Diff model provided the best quantitative and qualitative fits to the data, suggesting that it most closely mimics the human confidence computation.
January 26, 2026 at 7:18 PM
We found clear evidence that the Top2Diff model provided the best quantitative and qualitative fits to the data, suggesting that it most closely mimics the human confidence computation.
We then compared 7 confidence strategies: positive evidence (PE), Bayesian Confidence Hypothesis (BCH), Top-2 Difference in raw evidence (Top2Diff) or probability (ProbTop2Diff), Top Minus Average (ProbAvgRes), Entropy and Softmax. These are all the main competitors for multi-alternative decisions.
January 26, 2026 at 7:18 PM
We then compared 7 confidence strategies: positive evidence (PE), Bayesian Confidence Hypothesis (BCH), Top-2 Difference in raw evidence (Top2Diff) or probability (ProbTop2Diff), Top Minus Average (ProbAvgRes), Entropy and Softmax. These are all the main competitors for multi-alternative decisions.
Human subjects performed an 8-choice digit categorization task based on noisy MNIST images. We used RTNet - a network we developed recently that is known to show the signature of human perceptual decisions (Rafiei et al., 2024, Nat Hum Beh) - to model the internal activation produced by each image.
January 26, 2026 at 7:18 PM
Human subjects performed an 8-choice digit categorization task based on noisy MNIST images. We used RTNet - a network we developed recently that is known to show the signature of human perceptual decisions (Rafiei et al., 2024, Nat Hum Beh) - to model the internal activation produced by each image.
Congrats Chaz!!! Also, lovely kids :)
November 7, 2025 at 1:52 AM
Congrats Chaz!!! Also, lovely kids :)