Jeremy is headed "down under"

Jeremy has a long flight ahead of him. He will be travelling to Melbourna, Australia, in November to present his research at the International Conference on Neural Information Processing (ICONIP).

Two Opinions, One Action: Decoupling Extrinsic and Intrinsic Reward for Curiosity-Driven Reinforcement Learning
- J Zheng, W Yue, J Orchard
Curiosity is widely modeled in reinforcement learning as an intrinsic reward bonus added to the task’s extrinsic return. Such a `reward-shaping’ view conflates curiosity with reward maximization and thereby diverges from neuro-cognitive accounts in which curiosity and action selection arise from partially dissociable control processes. In this work, we introduce Decoupled Intrinsic–Extrinsic Control (DIEC), a basal-ganglia–inspired framework that treats curiosity as a separate control pathway rather than an additive reward term. DIEC maintains distinct intrinsic and extrinsic sub-agents and decides between them via an action-arbitration mechanism. Across benchmarks, DIEC substantially improves robustness and consistently increases sample efficiency. We attribute these gains to more flexible, context-sensitive regulation of the exploration–exploitation trade-off. Overall, revisiting how curiosity may be utilized by the brain, DIEC offers a simple, stable alternative to canonical intrinsic-reward shaping and a principled lens on curiosity as a separable control pathway interacting with goal-directed decision-making.

Well done, Jeremy and Wenhao! And bon voyage!