WhiteBox

A.I. for Impact

Deep learning for impact.

Monitor and assess environmental and social risks and impacts with deep learning. Optimisation and statistics are at the heart of machine learning. Fuelled by better data and supported by increasingly powerful hardware, machine learning is well suited to tackle challenging and complex problems, support decision-making and provide deep monitoring with high dimensionality. Artificial intelligence (A.I.) here means machine learning, and most often deep learning.

Inside a whitebox each dimension is a vector of hundreds of numbers, so a drought shows in all four at once.

The approaches

What the research built, and the rule it kept.

The programme’s work went into models that read across a whitebox’s dimensions at once. Representation learning, to turn mixed and unstructured sources, from interviews and disclosures to imagery and sensor streams, into the vectors a whitebox holds. Anomaly detection, to find the reading that does not fit, which is how fraud, waste and a site turning unsafe first show themselves. Forecasting, to say where a reading is heading before it gets there. And the fusion of all of it with what the people affected reported, so that a model’s output can be checked against the ground.

One rule held throughout: the models must show their working. Deep learning is powerful and opaque, which is why its models are called black boxes, and a black box inside a whitebox defeats the purpose. The research treated explainability as part of the model, not a report written about it afterwards.

Monitoring

Having the right frameworks, processes and procedures to collect quality and relevant data is the first step to being able to use A.I. to monitor and assess environmental, social and governance risks and opportunities for positive impact.

  1. Collecting and monitoring all data channels, internal and external to a matrix organisation.

  2. Creating new sources of quality and timely data.

Reporting

Once data collection mechanisms and frameworks are in place and environmental, social and governance risks and impacts can be assessed, timely reporting is critical.

  1. Reporting and taxonomy that aligns with standards and regulations.

  2. Reporting to all relevant stakeholders, both internal and external to the organisation.

In Druid, the product that came out of this work, both halves run every ten days on every field in the record: public satellites and field reports are the data channels, and every forecast is reported with its grade. How Druid does it