Module 4 - Governing Artificial Intelligence
Topic 1: External Governance
If certain laws and policies are applied depends on different aspects:
Where does the organisation that is developing or deploying an AI system operate?
Where are the users situated?
Where is the output used?
Depending on how these questions are answered, different laws have to be followed.
a. The Risk-Based Approach
- Minimal risk AI systems are allowed to be used freely. For example, spam filters or AI-
enabled video games are considered minimal risk.
- Limited risk systems have to adhere to transparency obligations. These regulations
concern for example chatbots that have to make transparent that they are bots and not
humans. Also, AI-generated content like text, audio or video has to be labelled as
artificially generated.
- High risk systems are described as tools that among others deal with critical
infrastructure, employment, essential private and public services, law enforcement that
may interfere with people’s fundamental rights or migration. It regulates how an AI
product must look like to be placed in the European market and put into service.
- Systems that fall in the unacceptable risk category are banned in the EU. This includes
for example social scoring systems, that rank people based on personal characteristics,
their behaviour or their social status.
b. Introduction into the AI Lifecycle
3 phases:
1. Design
2. Development
3. Deployment
* Design
- Problem formulation/ justification:
Justify why an AI application is the best solution for your problem (if there is a
cheaper, less risk prone, more effective and less complex approach to solving your
problem.)
Formulate the problem you are aiming to solve, and describe in concrete terms how
your AI system would solve the problem. You should decide how it would be
integrated in the non-technical project set-up, for example who will be using the
output of your AI system and in what way, and who is responsible for the use.
Concrete means here answering questions such as from which data you are deriving
which interpretations, or if your data source can actually answer your questions.
- Project planning:
Resource questions such as the budget of the project (including maintenance for
management and funding) or personal questions.
Drafting of policies and procedures: a data management plan, a data governance
plan, documentation and transparency measures such as risk assessment and legal
compliance.(it is advisable to include participatory processes in which the interests
of people affected by the system can be taken into account in the policies.)
- Data extraction/collection & Analysis:
The part of the AI Lifecycle regarding the data is one of the most fundamental steps in
the model and opens another smaller lifecycle: the data lifecycle. (planning,
acquisition, analysis, processing, publish/share.)
The acquisition of the data concerns where the data comes from
Analyse: is it suitable for the task at hand, are there any missing data points, is the
data in a good condition? This analysis will influence the following steps in the
development stage: which data preprocessing is necessary, and which model is
appropriate to be used with the data.
Sharing the data and making it accessible (if it is no sensitive data)
*Development
- Data Preprocessing:
Cleaning the data and adapting it to the project’s purpose.
Preprocessing may also involve decisions about what data to keep or to exclude -
decisions that can affect fairness and accuracy. For example, removing data on small
population groups because it is incomplete can lead to bias in how the model performs
for those groups.
- Model selection and training:
Have to consider what model design best suits our problem.
Datasets are split into training, test and validation datasets, for example in a 70%,
20%, 10% data split. For training, only the training dataset is used. It is important to
consider how the dataset is constructed and to distribute the data points in a way that
ensures that the training, test and validation sets all contain a similar distribution of
data points.
- Model testing and Validation:
At this stage, our trained model is tested to see how well it performs with unseen
data; i.e. how it would perform in a “real-life scenario”. Our validation dataset is
used to hone the model design and the hyperparameters.
Consider at which threshold (ngưỡng) we would consider our model to be successful.
What performance is necessary for our specific task? For example, a high risk task
would need a very high performance.
- Model Reporting: This step accompanies all the previous steps and involves documenting
what has been done.
* Deployment
- System implementation:
This is the stage after the model has been tested and finalised. Now the model can
be implemented: deployed and used by the target users.
Requires technical infrastructure, user training and communication with affected
communities.
- System use and Monitoring:
Monitor its usage and to keep testing the system’s performance.
Revisit some of the questions raised during the initial testing of the model: is the
model still fulfilling the original project’s goals? Does the actual use align with the
intended use? In other words, is the model being used by users for the purpose it was
designed for?
Monitoring should include feedback from users and the public => remains aligned
with its social purpose and does not quietly drift away from its original goals.
- Model udating: It’s important that any updates are documented, tested, and
communicated clearly to all users - especially if they influence how decisions are made.
Topic 3: Deep dive into Datasheets