Speaker

Shashank Srikant

Runs Manana Labs, a boutique consulting group specializing in AI and deep tech

Shashank Srikant has a PhD in computer science and runs Manana Labs, a boutique consulting group specializing in AI and deep tech. He is lead author of "Convergent Representations of Computer Programs in Human and Artificial Neural Networks" (NeurIPS 2022), which compares fMRI recordings of programmers with machine-learning models of code.

Research and background

Shashank Srikant has a PhD in computer science. His own website says he studied how programs are represented by both brains and machines, and that he runs Manana Labs, a boutique consulting group specializing in AI and deep tech. An MIT CSAIL spotlight describes him as working with Una-May O'Reilly and the ALFA Group. The ALFA Group's people page lists him among its PhD alumni (2023), and his own website lists machine learning, program analysis, and cognitive neuroscience as his research interests.

His episode covers Convergent Representations of Computer Programs in Human and Artificial Neural Networks, published at NeurIPS 2022 with Ben Lipkin, Anna Ivanova, Evelina Fedorenko, and Una-May O'Reilly. The study asks which aspects of a program the brain represents during comprehension. It uses fMRI recordings of programmers reading Python, first testing static and dynamic code properties such as syntax-tree and runtime metrics, and then training a probe to align brain activity with the representations of several machine learning models of code. The authors report that two brain systems, the Multiple Demand system and the language system, each encode specific properties of code and align with the machine-learned representations. They read this as evidence for at least two distinct neural mechanisms behind program comprehension, and suggest that code models may need training objectives beyond standard language modeling. Code and data are public on GitHub.

In Episode 005, Your Brain on Code, he reviews the cognitive neuroscience of reading code and shows how programs can be modified so that a model predicts stronger responses in particular brain regions. Lining up a model's internal representations with brain signals is an interpretability method: it shows what the model encodes and where it differs from people. The episode is filed under Cognitive Science, Neuroscience, and Interpretable AI.

Episode

References

  1. Shashank Srikant, Benjamin Lipkin, Anna Ivanova, Evelina Fedorenko, Una-May O'Reilly (2022). Convergent Representations of Computer Programs in Human and Artificial Neural Networks(opens in a new tab). NeurIPS 2022, Advances in Neural Information Processing Systems 35: 18834–18849. Discussed in Your Brain on Code: What Brains and AI Models Both Learn.

Browse every episode