Key Stages in VDML PROCESS

 By clearly defining the benefits the AI-enabled system is expected to deliver, we work with our clients to understand the context in which the system is being used. This sets a baseline for assessing risk. For example, applying AI in to assist with information retrieval and summarisation of large bodies of text carries a significantly different context to applying AI to supporting clinical diagnosis of x-rays. 

Based on the task, and the relevant legislation and industry compliance requirements that may apply, a risk assessment helps establish the required governance practices that need to be put in place. The risk assessment process includes understanding the data being processed by the system, and any data being used to develop or fine tune a machine learning model, and the likely impact of any errors the system may make. 

 Just as most organisations conduct User Acceptance Testing to validate a generic software solution has been configured appropriately to support their specific business processes, there is a need to validate machine learning models to ensure they will work within the intended context. 

Direct use of pre-built models or naive approaches to machine learning can lead to unreliable performance. It is typical for any machine learning component to be tested against the target data and fine-tuned if performance does not meet expectations. 

VDML provides a number of specific techniques by which KJR can help organisations test and tune machine learning components of their system to be confident they will perform as expected. By selecting testing data sets which are close to real world usage, and carrying out detailed error analysis, we can help our clients uncover underlying faults and limitations and put appropriate risk mitigations in place.

In situations where tuning has already been performed by a solution provider, independent validation of the model’s performance is key to a responsible approach to AI. 

 An AI-enabled system is much more than just a bare machine-learning model. ML components need to be integrated into a service delivery pipeline which enables appropriate interaction with the host organisation’s data and users. This can include ensuring ML components have access only to the appropriate data (e.g. ensuring client privacy is not being breached). Conversely, only the appropriate users have access to the ML components (e.g. to ensure models are not tampered with, or subject to unauthorised access). 

A key element of validation at this stage is ensuring any processes which are being used to enable the ML components to learn from and respond to live production data work as expected and have the appropriate governance controls in place. 

  By helping our clients to track the performance and integrity of their AI-enabled solutions from deployment, operation and maintenance, we can put in place the practical implementation of the governance controls identified as part of the risk assessment process. 

Being able to identify residual risks, detect model drift / sabotage, or simply measure performance and cost, the hosting organisation can identify opportunities for further optimisation and risk reduction.