Continuous Metalearning for AI lifecycle governance

Mind Foundry正在申请Innovate UK SMART Grant，为人工智能创新的一个关键领域--持续元学习提供资金。持续学习是人工智能模型的总称，这些模型在部署后继续学习和适应，不仅是关于它们的任务，而且关于它们自己的学习过程，使用他们定期收到的数据Mind Foundry声称，为了衡量、管理、解释和治理一个持续学习的人工智能系统，你需要运用元学习方法。这可能需要为特定人工智能系统中涉及的数据和模型创建护照和容器，捕获有关模型使用、性能和校准的必要数据和来源历史。一个不断发展的系统将需要积极的保障措施和规范，以确保其符合即将到来的监管要求，例如欧盟关于人工智能监管的新提案。为此，模型和系统在其整个生命周期中的演化完全公开和透明，以及对模型行为随时间变化的准确监控和管理，将成为公共和私营部门人工智能系统采购的先决条件。Mind Foundry方法的初始目标是不断检测人工智能系统的关键性能警告信号，例如目的漂移（已部署的模型以及数据以非预期的方式使用），模型过度拟合（通过过度使用单个训练数据集），数据治理（删除权），可解释性（解释权）和安全性（模型反转攻击的风险）。最终，这种方法可以扩展到人工智能系统，采取适当的缓解措施，并为人类用户发出警报并直接与人类用户合作。

Mind Foundry is applying for an Innovate UK SMART Grant to provide funding for a critical area of AI innovation -- continuous metalearning.Continuous learning is an umbrella term for AI models that continue to learn and adapt after the point of deployment, not only about their task, but also about their own learning process, using data that they receive periodically (through a continual process) or continuously.Mind Foundry asserts that in order to measure, manage, explain and govern a continuous learning AI system, you will need to apply a metalearning approach. This will likely require creating passports and containers for both data and models involved in a particular AI system, capturing necessary data and provenance history on the model's usage, performance and calibration over time.An evolving system will require proactive safeguards and specification in place to ensure it complies with upcoming regulatory requirements, such as the new EU proposal for AI regulation. To that end, full disclosure and transparency about model and system evolution throughout its lifecycle, as well as accurate monitoring and management of changes in the model's behaviour over time, will become a prerequisite for AI system procurement in both the public and private sectors.The initial goals of Mind Foundry's approach are to continuously detect the key performance warning signals of an AI system such as purpose drift (a deployed model, and thus data, being used in a way it was not intended), model overfitting (through overuse of a single training data set), data governance (right to deletion), interpretability (right to an explanation), and security (risk of model-inversion attack). Ultimately this approach can then be extended to AI systems actioning appropriate mitigations, and raising alerts for and collaborating directly with human users.