Cerebellum-inspired parallel deep learning

深度学习是一种机器学习，最近经历了戏剧性的发展。它已经对整个社会产生了影响，从科学发现到气候变化预测。然而，目前的深度学习模型需要数十亿个参数，可能需要数周的时间来训练，成本高达数百万英镑，碳足迹巨大。因此，为深度神经网络开发高效的训练方法变得越来越重要。深度神经网络效率低下的关键瓶颈之一是需要顺序执行大量计算步骤。在这里，受最近关于生物神经网络如何学习的发现的启发，我们建议为深度神经网络开发一种有效的训练算法。我们最近提出，一个专门的大脑区域，小脑，使大脑中的平行学习。小脑有两个特征可以很好地促进并行学习：稀疏性和模块性。首先，小脑包含高度稀疏的连接，每个神经元只有四个输入连接，这应该会导致更快的学习。第二，小脑是一个高度模块化的系统，能够很好地实现并行学习。受这些小脑特征的启发，我们将开发一个稀疏模块化系统，用于训练能够进行高效并行训练的深度学习网络。在与工业界的合作中，这种方法的好处将通过一种旨在加速机器学习的新型并行处理器来展示。总的来说，我们的工作将为并行深度学习带来一种新的方法，从而大幅减少训练时间和成本。

Deep learning, a type of machine learning, has recently undergone dramatic developments. It is already having an impact across society, from scientific discovery to climate change prediction. However, current deep learning models require billions of parameters which can take several weeks to train, costing millions of pounds with large carbon footprints. It is therefore becoming increasingly important to develop efficient training methods for deep neural networks.One of the key bottlenecks that underlie the inefficiency of deep neural networks is the need to perform a large number of computational steps sequentially. Here, inspired by recent findings on how biological neural networks learn, we propose to develop an efficient training algorithm for deep neural networks. We have recently proposed that a specialised brain region, the cerebellum, enables parallel learning in the brain. The cerebellum is defined by two features that are well placed to facilitate parallel learning: sparsity and modularity. First, the cerebellum contains highly sparse connectivity with only four input connections per neuron, which should result in faster learning. Second, the cerebellum is a highly modular system, which is well placed to enable parallel learning. Inspired by these cerebellar features we will develop a sparse-modular system for training deep learning networks capable of efficient parallel training. In collaboration with industry, the benefits of this approach will be demonstrated using a new type of parallel processor designed to accelerate machine learning. Overall, our work will lead to a novel approach to parallel deep learning, leading to a substantial reduction in training times and costs.