Model lineage / genealogy about substra-documentation HOT 5 CLOSED

The purpose of this feature is to create an identity card of a model that links it to datasets, algorithms, initial models.
This topic is related to the different works of Substra Foundation, both on the framework with composite training but also the topics trustworthy AI.
This notion of model genaology does not exist today, even if there are works in progress to be studied.
In particular DVC

from substra-documentation.

Outside of Substra Framework

The objective is to create a template that allows to list all the history of the predictive model.

DATA:

• How were they collected? Single dataset or assembly of several datasets?
• By whom?
• Is synthetic data used?
• Are personal or confidential data present?
• How the split between train, test and validation has been done?
• Use of PETs?
• How bias has been prevented?

ALGO:

• What is the learning algorithm used?
• By whom was the learning algorithm created?

MODELS:

• What is the Initial model?
• Performances of the model?
• Decision thresholds and ranges of indecision?
• Condition of use? Limits?
• How the model is used? How the model is reviewed over time?
• List of incidents

ACCOUNTABILITY:

• Who is responsible in case of incident? Human agency?
• What are the different levels of responsibilities between stakeholders?
• Use of subcontractors?

Substra Framework

How some of this elements could be collected in Substra framework? How Substra Framework can be used as element of proof?

from substra-documentation.

A very interesting paper on that matter: Model Cards for Model Reporting.

Also, Google's Model cards.

from substra-documentation.

Update from last MAP committee (10/09/2020):

• The objective here is to create an identity card of a model that links it to datasets, algorithms, initial models; that includes its “genealogy”!
• [Camille] an early early draft in the frontend already, sort of simple identity card. Not yet clean, readable, user-friendly. Could be great to work on improving that!
• The idea is to move forward in the different work groups on this ID card, and to see how to integrate it into the Substra Framework.
• [Amine] About monitoring data quality, and how it impacts model performance, continuous learning setups (or regularly re-trained models).
  • [Camille] This is rather done outside the framework today but we could probably imagine integrations with external tools via queries.
• Next step: the topic will be studied during the new season of DataForGood!

from substra-documentation.

Closing stale issue.

from substra-documentation.