Circuit mechanisms governing contextual feedback in vision
Motivated by design principles in biology
The brain has design principles that AI can learn from. Training on internet-scale data lets machines imitate behavior, but it rarely captures the underlying computations, making them brittle and hard to interpret. My work draws inspiration from neuroscience, incorporating principles that support robust, contextual behavior. Early on, I focused on modeling local and long-range lateral feedback pathways, producing models that differ qualitatively and quantitatively from black-box approaches and support a range of critical visual functions, including:
In my postdoctoral work, I am expanding this program with greater biological realism, incorporating top-down modulatory circuits that enable dynamic, context-dependent behavior (Govindarajan et al., 2024).
References
2024
NeurIPS
Computing a human-like reaction time metric from stable recurrent vision models
Lore Goetschalckx*, Lakshmi Narasimhan Govindarajan*, Alekh Karkada Ashok, Aarit Ahuja, David Sheinberg, and 1 more author
Advances in Neural Information Processing Systems, 2024
@article{goetschalckx2024computing,title={Computing a human-like reaction time metric from stable recurrent vision models},author={Goetschalckx, Lore and Govindarajan, Lakshmi Narasimhan and Karkada Ashok, Alekh and Ahuja, Aarit and Sheinberg, David and Serre, Thomas},journal={Advances in Neural Information Processing Systems},volume={36},year={2024},}
NeurIPS
Flexible context-driven sensory processing in dynamical vision models
Lakshmi Narasimhan Govindarajan, Abhiram Iyer, Valmiki Kothare, and Ila Fiete
Advances in Neural Information Processing Systems, 2024
@article{govindarajan2024flexible,title={Flexible context-driven sensory processing in dynamical vision models},author={Govindarajan, Lakshmi Narasimhan and Iyer, Abhiram and Kothare, Valmiki and Fiete, Ila},journal={Advances in Neural Information Processing Systems},volume={37},pages={3832--3856},year={2024},}
2022
NeurIPS (workshop)
The emergence of visual simulation in task-optimized recurrent neural networks
Alekh Karkada Ashok, Lakshmi Narasimhan Govindarajan, Drew Linsley, David Sheinberg, and Thomas Serre
@article{linsley2021tracking,title={Tracking without re-recognition in humans and machines},author={Linsley, Drew and Malik, Girik and Kim, Junkyung and Govindarajan, Lakshmi Narasimhan and Mingolla, Ennio and Serre, Thomas},journal={Advances in neural information processing systems},volume={34},pages={19473--19486},year={2021},}
2020
NeurIPS
Stable and expressive recurrent vision models
Drew Linsley*, Alekh Karkada Ashok*, Lakshmi Narasimhan Govindarajan*, Rex Liu, and Thomas Serre
Advances in Neural Information Processing Systems, 2020
The emergence of visual simulation in task-optimized recurrent neural networksIn SVRHM 2022 Workshop@ NeurIPS, 2022
Stable and expressive recurrent vision modelsAdvances in Neural Information Processing Systems, 2020