Explaining AI

From movie recommendations to medical diagnoses, people are increasingly comfortable with AI making recommendations, or even decisions. However, AI often inherits bias from the datasets that train it, so how do we know we can trust it? Dr. Harry Shum, Head of Microsoft’s AI and Research, breaks down some of the current biases in AI models. And then calls for us to open the "black box" in order to develop the transparency, fairness, and trust needed for continued AI adoption.

Highlights
The latest AI breakthroughs [0:24]
Xiaoice, the Chinese AI with EQ (as well as IQ) [2:42]
Why EQ leads to better digital assistants and chat bots [3:50]
How Japanese and Chinese businesses are using Xiaoice for sales and financial reports [4:51]
Gender bias in current AI models [6:22]
Mapping the gender bias with word pairings [8:33]
Harry Shum makes the case for transparent AI [12:21]
3 Reasons why we need explainable AI [12:58]
The tradeoff between accuracy and explainability in AI models [14:20]

Pull Quote
"...with IQ, we're helping people to accomplish tasks. And with EQ, we have empathy, we have the social skills, and the understanding of human beings' feelings and the emotions."

[SuperKillaki]

This topic is extremely important. More debates, discussion, presentations!

[swyxTV]

idk man, look, i'm as tech loving as the next a16z fan, but this rate of progress with xiaoice doesn't bother anyone else?

[kolmercm]

Great Video! anyone have a link to the graph from HKU Dr. Zhang's group, I had trouble finding it online. Link to the picture of AI recognition with years would be nice too!

[Rene-uzeb]

Word embeddings are useless. The concept that metric distance somehow can maintain all kinds of relationships is fundamentally flawed. Ever since BERT nobody uses them. The best an embedding can do is retain pairwise word distance, i.e. co-occurrence, which is trivial

[angelmcorrea]

Thanks a16z for share all this important topics .

[davidw]

Yes, man and women are not much different overall but have entirely different preferences and behaviours.

[tycn]

china: transparency transparency transparency (so we can train our AI). America/EU: Stop stealing my data!

[Rene-uzeb]

Explainable AI: makes about as much sense as those interview questions from companies like yours. Typical management thinking. Can you explain your thoughts? You have to ask the AI to explain itself obviously, instead of trying to untangle its neural circuits.

[gene]

I love that there is a conversation going on this important topic, not so happy about the good guy bad guys aspect, or as it was presented here Microsoft vs Google. This is a global problem, we should leave out the self advertising from this debates.