Project Selection
This section sets out the ethical consideration process for Branch Leads to follow when selecting projects to take on for the next project cycle.
Each Branch will have its own specific needs and existing approaches to what is required for a project to be selected. This part of the Framework is designed to supplement these things by building in consideration of ethical implications.
To maximise effectiveness, the ethical consideration element of the project selection process is applied once the following have been established for a given project:
- Project and aim – what is it that you are proposing to create?
- Defined scope – what area will you be working in? What is the end goal?
- Proposed tangible outcome – what will be the result of this project?
- Estimated project length – will it run for one project season, or multiple?
In order for a proposed project to be eligible for selection, it must be demonstrated that it will promote and preserve the following principles:
| Respect for human agency | Privacy, personal data protection and data governance | Fairness | Individual, social, and environmental well-being | Transparency | Accountability and oversight |
|---|---|---|---|---|---|
| Human beings must be respected to make their own decisions and carry out their own actions. Respect for human agency encapsulates three more specific principles, which define fundamental human rights: autonomy, dignity and freedom. | People have the right to privacy and data protection and these should be respected at all times. | People should be given equal rights and opportunities and should not be advantaged or disadvantaged undeservedly. | AI systems should contribute to, and not harm, individual, social and environmental wellbeing. | The purpose, inputs and operations of AI programs should be knowable and understandable to its stakeholders. | Humans should be able to understand, supervise and control the design and operation of AI based systems, and the actors involved in their development or operation should take responsibility for the way that these applications function and for the resulting consequences. |
This is demonstrated by a Statement of Compatibility that is prepared by the Branch Lead in consultation with a member of the Law & Ethics Committee. This is a document designed to ensure a proposed project has been reviewed on the basis of its ethical implications before it is selected to proceed to the design phase.
CASE STUDY: ‘CT Facial Reconstruction’
CT Facial Reconstruction was a past Deep Learning project at MDN. The overall aim of the project was to create, implement and train a set of neural network models that accurately generated reconstructed generalised 3D volume images from 3D CT head scans. This model was intended for use in the field of forensic facial reconstruction.
Examining the project overview, there are a number of ethical concerns that may arise. The following principles have been selected to highlight the key concerns.
Privacy, personal data protection and data governance: The data used to train the neural network models included up to 5000 scans from drug overdose victims. This raises concerns of privacy and data protection. It may be unclear whether these deceased individuals had provided consent for their images to be viewed and used for ‘proof of concept’ and training purposes.
Fairness: The model is being trained on large data taken from individuals of caucasian descent. This raises the question of fairness and data bias, as the model may be skewed to recognise and function with identifying caucasian features. This leaves identifying individuals from non-caucasion backgrounds a difficult task. It is noted that the MDN model has been created for conceptual use, instead of practical forensic application.
Reliability and safety: Alongside fairness, the issue of reliability can also arise. The project acknowledged the instability of adversial learning, and the need for other training techniques that will improve model stability and hyperparameter sensitivity. It is important that in real life use, these models have sufficient monitoring and testing to ensure reliability and accuracy.