GDPR and AI – strategies and options in a nutshell

GDPR and AI – strategies and options in a nutshell

I had the pleasure to meet @KirkDBorne last week and participate in a panel moderated by Kirk – Decoding #GDPR, #IoT, and #UX in the #BigData world: discussing #AI #MachineLearning #DataSecurity at the @BigDataWorld_  / Security of Things organized by @closerstill1. On the panel were also @maratsoumari @clurr
I was discussing GDPR and AI. This post presents my personal views. Note I could not discuss all of the below due to time constraints. I also conduct a new course on AI/Deep Learning applications where I discuss these strategies
GDPR essentially consists of four things

Explicit consent.
Right to be forgotten.
Data portability.
Algorithm transparency

Depending on how its interpreted, the last item may make AI potentially illegal because it includes elements such as
         a) Providing meaningful information about the logic involved to customers
         b) Explaining significance and consequences of the choice / strategies/ algorithm
         c) Right to human intervention on a decision if requested
 
So, what could we do about it?

        a) We could watch the DARPA challenge for Explainable AI for any insights
        b) The LIME framework could provide a strategy. This works by injecting noise and observe what features are
            driving the score. The LIME framework paper(pdf) claims to explain predictions of any classifier in an
             interpretable manner by learning an interpretable model locally around the prediction.
        c) One more option is we could deploy models that are built to express interpretability on top of inputs of the AI
             model. Cristoph Molnar’s free book on interpretable ml explains this well
        d) Finally, In the log run – we could adopt a strategy beyond Back propagation as Geoffrey Hinton seems to be
             alluding recently Geoffrey Hinton believes we should rethink backpropagation
Have I missed any strategies?
On a more practical front, I think businesses will take the approach of preparing and documenting metadata including the inputs, outputs and the model to provide a degree of explainibility
I discuss these ideas in the new course on AI/Deep Learning applications  
Image source ico
 

Link: GDPR and AI – strategies and options in a nutshell