UC Santa Cruz

Understanding how biological neural networks learn is both a fundamental scientific question and an increasingly urgent practical one. Despite decades of work demonstrating that in vitro neural cultures can be modified through electrical stimulation, the field lacks both a shared experimental framework for biological learning, as well as results mirroring those of artificial neural networks. This dissertation presents the tools, experiments, and scientific findings developed while pursuing biological learning in a dish. The work begins with a hybrid soft-rigid robotic system where model-free reinforcement learning outperforms classical control, establishing a first learning environment framework and motivating the extension of reinforcement learning from artificial to biological substrates. Contributions to closed-loop optogenetic modulation of epileptiform activity in human brain slices and a cloud-connected platform for organoid-based neuroscience education provided foundational infrastructure and early validation that high-density microelectrode arrays could reliably interface with biological tissue for stimulation-driven experiments. The central result is the first demonstration of goal-directed learning in brain organoids. Cortical organoids interfaced with high-density microelectrode arrays learned to balance an inverted pendulum (the ``Cartpole'' task) through closed-loop electrophysiology. Training signals were adaptively selected by a reinforcement learning algorithm and significantly outperformed both random stimulation and unstimulated controls. Pharmacological blockade of glutamatergic transmission abolished the learned improvements and washout restored them, confirming that the adaptation is biologically mediated and requires synaptic transmission. Causal connectivity analysis revealed that the strength of direct stimulus-evoked pathways predicts learning performance greater than traditional measures of functional connectivity, providing a basis for selecting electrode configurations prior to training. Two final technical contributions provide a foundation for future exploration: RT-Sort, a real-time spike sorting algorithm, and BrainDance, an open-source Python framework for designing, running, and analyzing closed-loop electrophysiology experiments. BrainDance provides a modular architecture capable of interfacing with various hardware devices. By design, BrainDance facilitates accessible design and sharing of experiments. Together, these tools bring rapid-iteration to biological neural learning, allowing any laboratory with compatible hardware to run reproducible experiments on their own biology.