Page 13 - AI Ready – Analysis Towards a Standardized Readiness Framework

AI Ready – Analysis Towards a Standardized Readiness Framework

As an example, an advanced driving assistance system (see clause 4.2.3) involves different
car manufacturers with different implementations who might adopt different parameters, the
divergence in implementation might create lock-in situations for users preventing flexibility
and choice of vendors. Additionally, issues concerning data privacy, data protection, and
responsibilities are to be studied collaboratively in open standards such as those developed
by ITU, which will ensure secure, trustable, and interoperable end-to-end solutions.

5) Developer Ecosystem created via Opensource

Cloud-hosted solutions with exposed APIs for subscribing/publishing data from portals [49]
would create value for the overall industry and lead to innovative applications that solve real-
world problems using AI/ML. A prime example is research solutions for satellite data usage in
the fire propagation model [51].

Reference solutions, open models, and toolsets created in opensource help in mobilizing
research and innovation, acting as a baseline for AI integration, which could be extended,
enhanced or optimized based on specific use case requirements. Solutions published as a
result of ITU AI/ML Challenges such as the TinyML Challenge [66] are good examples of open,
published, and developer-driven solutions.

6) Data collection and model validation via Sandbox pilot experimental setups

ITU defined ML Sandbox in [ITU-T Y.3172] and described the details of Sandbox architectures
in [ITU-T Y.3181]. In essence, Sandbox is an environment in which machine learning models
can be trained and their effects tested and evaluated before deploying in the real world. This
has since seen wider applications in various use cases.

Implementing continuous improvement of models using feedback and optimizations in the
Sandbox helps to optimize essential tasks within disaster-stricken areas [52]. Unmanned aerial
vehicles (UAVs) can learn and adjust their operations (including route navigation, returning
to charging stations, and data detection and transmission) based on feedback from the
environment.

For example, traffic regulation scenarios using visual cameras [36] and other sensors use AI/
ML feedback loops, which collect data, produce inferences, create action recommendations
and policy applications, and are tested and validated using pre-built traffic plans for specific
occasions.

Pilot setups via Sandboxes can help in assimilating local communities and utilities into the
solution. For example, in [51], fire detection and propagation models are tested and validated,
and alarms are used to provide advanced information to local communities and utilities.

6