Ajay Subramanian

I use psychophysics methods to probe the differences underlying human and machine vision. I am currently studying the computations that make human object recognition more robust than neural network recognition, and how these computations could help us develop more generalizable computer vision systems. I am a PhD student in Cognition & Perception at New York University, advised by Denis Pelli. Previously, I completed my undergraduate degree in Electronics and Communication Engineering at BITS Pilani, India. My broad interests are in deep learning inspired by and applied to human psychology and neuroscience.

• Subramanian, A., Sizikova, E., Majaj, N. J., Pelli, D. G. (submitted, 2023). Spatial-frequency channels, shape bias, and adversarial robustness.
• Zhou, J. Y., Chun, C., Subramanian, A., Simoncelli, E. P. (submitted, 2023). Comparing models of neural representation based on their metric tensors.
• Tsividis, P., Loula, J., Burga, J., Rodriguez, J. P., Arnaud, S., Foss, N., Campero, A., Subramanian, A., Pouncy, T., Gershman, S., Tenenbaum, J. B. (submitted to Nature, 2023). Human-Level Reinforcement Learning through Theory-Based Modeling, Exploration, and Planning.
• Subramanian, A., Price, S., Sizikova, E., Kumbhar, O., Majaj, N. J., Pelli, D. G. (arXiv, 2022). SATBench: Benchmarking the speed-accuracy tradeoff in object recognition by humans and dynamic neural networks. arXiv.

• Subramanian, A., Chitlangia, S., Baths, V. (2022). Reinforcement learning and its connections with neuroscience and psychology. Neural Networks, 145, 271-287.
• Turner, J. P., Knight, J. C., Subramanian, A., Nowotny, T. (2022). mlGeNN: accelerating SNN inference using GPU-enabled neural networks. Neuromorphic Computing and Engineering, 2(2), 024002.

• Subramanian, A., Sizikova, E., Majaj, N. J., Pelli, D. G. (2023). Spatial-frequency channels for object recognition by neural networks are twice as wide as those of humans. An explanation for shape bias? ECVP, Paphos, Cyprus.
• Subramanian, A., Sizikova, E., Majaj, N. J., Pelli, D. G. (2023). Spatial-frequency channels for object recognition by neural networks are twice as wide as those of humans. VSS Meeting, St. Pete Beach, USA.
• Zhou, J. Y., Chun, C., Subramanian, A., Simoncelli, E., P. (2023). Computing and comparing metric tensors in neural response models. VSS Meeting, St. Pete Beach, USA.
• Subramanian, A., Sizikova, E., Majaj, N. J., Pelli, D. G. (2023). Spatial-frequency channels for object recognition by neural networks are twice as wide as those of humans. COSYNE, Montreal, Canada.
• Subramanian, A., Price, S., Sizikova, E., Kumbhar, O., Majaj, N., Pelli, D. G. (2022). Benchmarking dynamic neural-network models of the human speed-accuracy tradeoff. VSS Meeting, St. Pete Beach, USA.
• Subramanian, A., Patil, R., Baths, V. (2022). Word2Brain2Image: Visual Reconstruction from Spoken Word Representations. ACCS, Goa, India.