Adaptive Neural Networks through Epigenetic Processes

最近数据和计算的可用性增加导致了全球对人工神经网络的兴趣--人工神经网络已被证明在创建人工智能方面非常有效，特别是在大规模(例如，深度)运作时。部署神经网络通常涉及训练，以在输入和输出之间建立适当的权重。然而，这种培训取决于情景，如果基础环境发生重大变化，则需要进行再培训。因此，对于未来的神经网络来说，非常希望能够根据其部署场景的变化而自适应，以便在不进行全面再训练的情况下具有明显的灵活性和保持性能。随着人工智能与动态环境的接触，这一点将变得更加重要。在动态环境中，神经网络可能需要以更大的灵活性进行理想的学习和适应。在这个项目中，目的是从遗传算法(GAs)中获得神经网络自适应的灵感。长期的工作已经证明，遗传算法可以成功地用于进化高性能的神经网络群体(例如，通过扩大拓扑来进化神经网络-NEAT)，而优步人工智能实验室最近的工作已经证明，遗传算法是一种非常有竞争力的替代方案，可以训练用于强化学习的深度神经网络。该项目将涉及表观遗传算法这一相对较新的领域的工作，并利用这些算法来促进以遗传形式表示的神经网络的适应。遗传过程启发了表观遗传算法，在这种算法中，与环境的相互作用会影响个体在种群繁殖周期之间持有的基因。表观遗传学是生物学、药理学和医学中公认的一种现象，有许多应用。该方法将包括开发合适的表示法、表观遗传学技术和评估方法，以探索神经网络中的自适应问题。这是一个雄心勃勃的目标，该项目将寻求通过评估众所周知的基准，在这一领域确立基本步骤。该博士学位适合热衷于深入了解最先进的深度学习和神经进化，并对开发具有新的通用能力的未来人工智能感兴趣的人。该项目将与IBM UK、DSTL、加的夫大学犯罪和安全研究所以及国际合作(通过分布式分析和信息科学分布式技术联盟在美国的大学合作伙伴)合作进行。

The recent increased availability of both data and computation has led to global interest in artificial neural networks - these have been shown to be highly effective in creating artificial intelligence particularly when functioning at scale (e.g., depth). Deploying neural networks typically involves training to establish appropriate weights between inputs and outputs. However such training is scenario dependent, with retraining needed if there is a significant change in the underlying environment. Therefore it is highly desirable for future neural networks to have the ability to self-adapt in relation to changes in their deployment scenario, so that flexibility is apparent and performance is maintained without full-scale retraining. This will become more important as AI engages with dynamic situations where neural networks may need to ideally learn and adapt with greater agility. In this project the aim is to take inspiration from genetic algorithms (GAs) for neural network self-adaptation. Longstanding work has demonstrated that GAs can successfully be used to evolve a high performance population of neural networks (e.g., Evolving Neural Networks through Augmenting Topologies - NEAT) while more recent work by Uber AI labs has demonstrated that genetic algorithms are a remarkably competitive alternative to training deep neural networks for reinforcement learning. The project will involve working in the relatively new field of epigenetic algorithms, and using these to promote adaptation of neural networks that are represented in a genetic form. Genetic processes inspire epigenetic algorithms, where the interaction with the environment influences the genes held by an individual between cycles of population reproduction. Epigenetics is a phenomenon that is now recognised in biology, pharmacology and medicine with many applications. The methodology will involve developing suitable representations, epigenetic techniques and evaluation methods to explore the problem of self-adaption in neural networks. This is an ambitious goal and this project will seek to establish fundamental steps in this area by assessing benchmarks that are well understood. This PhD is suitable for someone keen to gain in-depth knowledge of state-of-the-art deep learning and neuroevolution, and an interest in developing future AI with new and general capabilities. The project will be carried out in collaboration with IBM UK, Dstl, Cardiff University Crime and Security Institute and with international cooperation (University partners in the US via the Distributed Analytics and Information Sciences Distributed Technology Alliance).