Responsible AI support

[SergeStinckwich]

Dear all,

I see that responsible AI is nicely presented in the chapter 17, as an engineering constraint, with topics related to explainability and fairness. Are they any plans to add some software support in order to have a framework that is more integrated with tools like TinyTorch to assess the different kinds of fairness (group or individual fairness) or explainability methods like SHAP? Or the idea is to use external tools like FairLearn.

Best regards,

[Shashank-Tripathi-07]

Hi, @profvjreddi would help the best on this, tagging him so that it reaches his feed and get's him here first :)

regards,
Rocky

[profvjreddi]

Hi folks, sorry for the late reply. I've been swamped getting some other research-oriented work going (https://arch2.mlsysbook.ai/). Anyway, with respect to this...

Great question, Serge. The book's software support tackles responsible AI at the systems level rather than at the algorithm level. Each chapter ships with a hands-on lab, and the responsible AI labs treat fairness, accountability, and monitoring as measurable engineering commitments (what fairness monitoring costs at serving time, who owns the threshold, what makes a dashboard actually accountable) rather than reimplementing the assessment methods themselves. The labs are still very much under development, though, so feedback and ideas like this are exactly what helps us figure out what belongs where.

For the assessment methods you mention (group vs. individual fairness, SHAP-style explainability), the intent is exactly what you guessed. Lean on established external tools like Fairlearn and SHAP rather than build a parallel framework. TinyTorch is deliberately scoped to building framework internals (tensors, autograd, training loops) so students see how ML systems work under the hood. Fairness toolkits are mature ecosystems, and reimplementing them would teach less than wiring them in.

That said, a worked example or lab that plugs Fairlearn or SHAP into a training/eval loop and connects it back to the chapter's framing would be a genuinely nice addition. If you'd be interested in sketching that out, I'd be happy to work with you on it.

[SergeStinckwich]

hi Vijay. Thank you for the feedback.
I'm definitely interested to build a lab that plugs Fairlearn as you mentioned. I recently have done a Master lecture "From Data to Policy: Interpretation, Communication, and Ethics" where I discussed fairness metrics theory and I want to have a more practical exercice for the next session.

[profvjreddi]

Oh that's awesome. If there is anything I can read, please share. Happy to dig into it.

[SergeStinckwich]

I have slides but there is nothing fancy, just basic definitions of fairness metrics (demographic parity, equalised odds, ...), similar to what you have already in chapter 17. I will think about a specific use case to use FairLearn and I can share then.