Why do you care about AI Existential Safety?

With the increasing capabilities of AI systems, it will become imperative to understand as well as ensure that these models behave in ways that are in accordance with human values and preferences. Such alignment, although possible to be achieved implicitly during learning, will most likely have to be injected via explicit mechanisms. If we fail to achieve such alignment, this can result in catastrophic outcomes, including existential risk for humanity.

Please give at least one example of your research interests related to AI existential safety:

• Empirical theory of deep learning i.e., trying to understand deep learning systems in a better way, before trying to maximize their performance
• Interpretability i.e., try to develop models that can explain their predictions in a faithful manner
• Develop systems that are more biologically plausible in order to better understand how such values can be hard-coded into the model directly, without asking them to learn it from data. Humans seem to have a moral compass that guides them independently of their world knowledge.