The question under all of it: how can embodied or resource-constrained agents build predictive internal models of the world, adapt continuously after deployment, and learn efficiently using local, predictive, or non-standard learning rules?
A spiking version of CLP, co-designed with neuromorphic silicon, that learns new classes from a data stream on the chip itself, without replay buffers or backpropagation.
An end-to-end on-chip pipeline that learns new human actions from an event-camera stream, class-incrementally, without sending data off the device.
A prototype-based online learner for robots that meet new objects after deployment: it learns from few labels, detects novel objects and keeps learning them without supervision, with no replay buffer.
A spiking network on Intel Loihi that learns new objects in a short interactive session with a user, with a neural state machine deciding on-chip when and how to learn.
PCN-TA, a predictive-coding online learner that keeps latent states across consecutive frames of a sensory stream, cutting the cost that usually makes predictive coding expensive.
A spiking adaptive controller inspired by the cerebellum, running on Intel's Loihi chip and controlling a simulated two-joint robot arm in closed loop, in real time.
Fine-tuning the released SmolVLA LIBERO policy on new LIBERO-90 manipulation tasks in simulation and measuring how well its earlier skills are retained across continual-learning baselines. Planned extension: how low-bit compression changes what a VLA keeps under fine-tuning.
A compact, self-supervised predictive world model with continual learning across multiple time-scales, evaluated on the LIBERO manipulation benchmark. The question I'm chasing: can an agent keep adapting (at test time, across tasks, over its lifetime) instead of freezing the moment it ships?
Interactive continual learning for robots
2022
Master's studies, German Academic Exchange Service
Computational Neuroscience
