Dr. Ravi Kiran Sarvadevabhatla gave an inaugural day talk on Deep Neural Networks for Part-based Object and Scene Representations at a workshop on Brain, Computation, and Learning (BCL-2023) held at Indian Institute of Science, Bangalore (IISc) from 9 – 13 January.
Here is the summary of the talk in Dr. Ravi Kiran Sarvadevabhatla’s words:
Aristotle is said to have famously remarked, ‘The whole is greater than the sum of its parts’. Philosophy aside, identifying semantically meaningful parts of an object can enable a richer understanding of visual content. In this talk, I present deep neural networks for part-based understanding of objects and scene representations. In its original form, the task setting naturally induces a combinatorial explosion of deep network outputs. As part of the talk, I show that the inductive bias of the task can be exploited to craft deep network architectures which mitigate the combinatorial issue. The capabilities of the resulting deep networks suggest a potential for applications in cognitive studies, graphics and robotics.
BCL-2023 homepage: https://bcl.iisc.ac.in/