CapeTownAI

Global network AI practitioners in 40+ cities who share challenges and lessons learned in applied AI. Join our community & meet at our 100+ peer2peer gatherings

[CapeTownAI @CapeTownAI]

The responsibility falls on the company or individual to explain the discrimination. Good reasons are under section 14 of Pepuda Act @Danielle_Louw_ #CriminalJustice

[CapeTownAI @CapeTownAI]

"For AI, we have to check listed ground such as marriage, gender, race etc, but also check against all permutations of listed ground" @Danielle_Louw_ #CriminalJustice

[CapeTownAI @CapeTownAI]

South Africa has legislation on direct and indirect discrimination @Danielle_Louw_

[CapeTownAI @CapeTownAI]

"Maths needs morals and science needs a conscious " @CummingsRenee #CriminalJustice

[CapeTownAI @CapeTownAI]

The four questions we need to ask: Who’s feeling the algorithms? what are we feeding them? how is the data selected and who is selecting the data? @CummingsRenee #CriminalJustice

[CapeTownAI @CapeTownAI]

how easily can we dupe facial recognition. Google mistook the face of a cat for guacamole @CummingsRenee #CriminalJustice

[CapeTownAI @CapeTownAI]

How raw is your raw data? We got to look at where our data comes from. Police data is some of the dirtiest because it lacks transparency and accountability. @CummingsRenee #CriminalJustice

[CapeTownAI @CapeTownAI]

White collar crime high risk area are 90.12% accurate in seeing where the financial crime will happen @CummingsRenee #CriminalJustice

[CapeTownAI @CapeTownAI]

The crucial question of whether they help judges make better decisions are unanswered. The algorithms are developed with the best intentions but aren’t working that way @CummingsRenee #CriminalJustice

[CapeTownAI @CapeTownAI]

Criminal justice is a great case study because the stakes are high and we should look carefully at this @CummingsRenee #CriminalJustice

[CapeTownAI @CapeTownAI]

@CummingsRenee echoes the question by Ruha Benjamin "are we perpetuating deeply discriminating practices under the guise of reason, neutrality and social improvement?"
#CriminalJustice

[CapeTownAI @CapeTownAI]

We need to ask more questions before purchasing and using the tools. Algorithms are only good as the data they work with and many have inherited the wide spread biases in our society. We are automating inequality. @CummingsRenee

[CapeTownAI @CapeTownAI]

It is encoding systemic baises into the algorithm. We are seeing a similar challenge in the UK. The history of the US

[CapeTownAI @CapeTownAI]

Arraignment by algorithms = everything in criminal justice is done by risk scores and is done with opaque algorithms and zombie predictions which over estimate risks @CummingsRenee

[CapeTownAI @CapeTownAI]

It’s great to be at @ACCapeTown with @CummingsRenee speaking on AI and criminal justice #AmericanCorner #CriminalJustice

[Chantal Lailvaux @chantallailvaux]

Join ⁦us to hear international speaker Renée Cummings on
*Inclusion and Diversity in AI: Criminal Justice* with Neil du Toit. 2 October. Get your tickets below. ⁦@CapeTownAI⁩ ⁦@ACCapeTown⁩ #AI #AIlaw #AIethics https://t.co/k54zBZESf7

[CapeTownAI @CapeTownAI]

There are signs for the precursor to #singularity but it’s impossible to say when it will happen comments #SebastianBodenstein and then adds that there at vast global resources committed to this at the moment #MachineLearning #AI

[CapeTownAI @CapeTownAI]

#SebastianBodenstein if you want to find a new method, it’s hard because there isn’t the hardware for it. This makes it challenging to find alternatives to #backprop #MachineLearning #AI

[CapeTownAI @CapeTownAI]

#SebastianBodenstein notes that humans have evolved over 500 millions years. It is not that each of us has learnt over a few months or years. We draw from evolutionary knowledge #MachineLearning #AI

[CapeTownAI @CapeTownAI]

Real nets use spikes rather than real activations
Deals Law
No weight-tying

[CapeTownAI @CapeTownAI]

#SebastianBodenstein information flows one way along a neuron #MachineLearning #AI

[CapeTownAI @CapeTownAI]

For backprop, the feedforward weights must be the same as the feedback, so we would need a subnetwork. We know this can’t happen because of the weight transport problem, but that’s not all!