Development and Commercialisation of AI certification platform for the banking sector

在COVID-19之后，各种规模的企业都需要贷款来生存、重建和恢复经济成功。银行将需要向经济释放流动性，做出正确的贷款决定，而不偏向或歧视经济和社会的任何部分。与此同时，由于COVID-19，这些银行正在经历严重和意外的贷款减值，因此，他们对信贷采取了谨慎的态度。现代银行已经利用机器学习系统快速有效地评估其贷款业务中的贷款申请并采取行动。然而，银行目前依赖现有的合规和治理结构来管理这些新的决策系统，这些结构缺乏知识和专业知识来预测机器学习系统可能出现的故障。银行自己预测，他们将越来越依赖于他们的ML模型可以提供的自动化和可扩展性。这对确保银行遵守行业标准和法规（如非歧视或数据隐私）构成了重大风险。通过这个为期9个月的行业研究项目，Mind Foundry Ltd将提供一个工具，解决ML模型的合规性问题，以批准贷款申请。该解决方案将包括一个认证系统，其中包括贷款审批领域ML模型的合规性定义，当经过训练的ML模型不符合时暴露的方法，以及将这些方法应用于经过训练的模型的机制。

In the wake of COVID-19, businesses of all sizes will need loans to survive, rebuild, and grow their way back to economic success. Banks will be needed to release liquidity into the economy, making the right lending decisions without being biased towards or discriminating against any sections of the economy and our society. At the same time, these same banks are experiencing severe and unexpected loan impairment due to COVID-19, and, as a result, they are taking a cautious approach to credit.Modern banks already make use of Machine Learning systems to quickly and efficiently evaluate and act on loan applications across their lending operations. However, banks currently rely on existing compliance and governance structures to manage these new decision-making systems, structures which lack the knowledge and expertise to anticipate the possible failings of Machine Learning systems . Banks themselves project that they will become increasingly reliant on the automation and scalability that their ML models can and do provide. This poses a significant risk in ensuring that the banks comply with industry standards and regulations, such as non-discrimination or data privacy.Through this 9 month industrial research project, we at Mind Foundry Ltd will deliver a tool addressing compliance of ML models for approving loan applications. The solution will consist of a certification system comprising a definition of compliance for ML models in the field of loan approval, methods for exposing when trained ML models are not compliant, and a mechanism for applying these methods to trained models.