Meta-learning from a dual machine-learning and biological perspective

对于机器学习和神经科学来说，深度网络学习仍然是一个巨大的挑战。在ML方面，人工神经网络的训练通常很慢，需要数百万个训练样本，而且往往依赖于广泛的超参数优化。在神经科学方面，大脑似乎最傲慢地处理所有这些问题--然而支撑学习的神经机制仍然难以捉摸。目前还不清楚人工深度学习算法是否以及如何在大脑回路中实现。这个项目将通过研究元学习将ML和神经学习方法统一起来，即学习策略的学习(I)克服现有算法的限制，(Ii)更接近大脑使用的生物物理可行解决方案的类型。这项工作将包括数学理论和数值模拟。

Learning in deep networks remains a significant challenge for both machinelearning and neuroscience. On the ML side, training artificial neural networks is generally slow,requires millions of training examples, and often relies on extensive hyper-parameteroptimization. On the neuroscience side, brains appear to navigate all these problemsalmost insolently -- yet the neural mechanisms that underpin learning have remainedelusive. It is unclear whether and how artificial deep learning algorithmscould be implemented in brain circuits. This project will unify ML and neuro approaches tolearning by studying meta-learning, i.e., the learning of learning strategies that (i)overcome the limitations of off-the-shelf algorithms and (ii) draw closer to the type ofbiophysically feasible solutions which brains use. The work will consist of mathematicaltheory and numerical simulations.