Will YOU trust the algorithm?

EconTalk Extra
by Amy Willis
Pedro Domingos on Machine Lear... Leif Wenar on Blood Oil...

EconTalk host Russ Roberts shows a pessimistic side this week in his conversation with Pedro Domingos on machine learning. What promise does the future of machinae learning hold? What's the difference between AI and machine learning? Will a Master Algorithm become possible when the best of these schools of machine learning are joined? And more interestingly, will you trust it?

These were just some of the issues touched on in this fascinating conversation, on which we'd like to hear your input as well. Consider some of the questions posed here, and share your response in the Comments. And/or try some of these thought experiments out on your friends...You just never know where a great conversation might take you...

robot baby.jpg

1. We're somewhat accustomed to Russ's optimism about all things technology, but this week was a little different. How would you characterize the nature of Russ' skepticism about the potential for the emergence of a Master Algorithm? To what extent do you share his skepticism?

2. What would be the purpose of the creation of baby Robby the Robot? How would Robby purportedly teach us about how we learn, and how much potential do you think there is in this approach?

3. I've been saying for a long time that amazon knows me better than my husband. Anybody else feel that way? How comfortable are you with the notion of having a "digital avatar, " as Domingos suggests? How might the problem of avatar ownership be solved? (That is, we likely all agree with Domingos that we wouldn't want Google to own our avatar...)

4. Domingos is confident that not many jobs will be lost immediately due to machine learning. But he also admits it may be possible someday to live in a world where very few humans need to work. Do you agree with his view that in such a world a universal basic income would make sense? Why or why not?

Comments and Sharing

TWITTER: Follow Russ Roberts @EconTalker

COMMENTS (4 to date)
Nicolas writes:

I am currently enrolled in an economics PhD program. I have been listening to econtalk since my first year. I really enjoy the podcast, but I have to say I was very disappointed by this episode. I was looking forwad to hearing how Professor Domingos would distinguish beween econometrics and machine learning, and came away unsatisfied.

Professor Domingos seems to believe that machine learning has a monopoly on non-linear methods. This is a ridiculously ignorant statement to make. I'm surprised that a professor of machine learning at a prestigious university is apparently unaware not only of non-linear least squares, non-parametric methods such as Kernel estimation and so on, or even the humble idea (taught in first year undergraduate econometrics classes) that you can put non-linear terms on the RHS of a multiple linear regression. Note that this is partially analogous to the Kernel trick popular in machine learning.

The entire emphasis of the podcast - that machine learning is something fundamentally new where a machine programs itself and so on - is exaggerated. By this definition, when I put into stata "reg y x", the computer is "learning" how to predict y given x. This is just regular old univariate linear regression.

There are incredible things that ML can do that econometrics can't, or at least doesn't due to its emphasis on causal interpretations. ML seems to have a set of very useful tool for quasi-nonparametric and, in particular, forecasting. But this conversation was at a very low level, and Russ didn't even defend Domingos' painfully ignorant statements about econometrics.

Warren writes:

I may be a little confused about the terms so let me specify them.
Machine learning is running a set of algorithms as we do now. A Master Algorithm is a set of algorithms that generalizes: this being comparable to an computer with an operating system only. Because the computer is a 'general' it can 'learn' whatever algorithm you feed it.
Yes, I think we will probably have a Master Algorithm some day that learns.

Will we have a singularity, meaning a computer that is conscious and learns in the same way a human does. NO
Our scientists still have no idea how consciousness comes about; of course there are many ideas. And intelligence requires emotions, spirits, interest, etc.

Mike Wengler writes:

Regarding the question, will a machine ever have consciousness, I will comment. Domingos said if I recall correctly, a machine will get to the point where it SEEMS conscious to us, as conscious as we seem to each other. But we can never really know if it FEELS conscious to itself, just as we never actually know about each other if we each FEEL conscious, or if it is only me.

I'd like to point out that to the extent you believe you feel conscious, you are carrying around a working example of a machine that feels consciousness: your brain. It is not a machine made of transistors and gears, but of neurons and glial cells. But it is a machine, in the sense that as far as we can tell, each component microscopically follows the same "simple" physical laws as do the gears and transistors in the machines we build. So regardless of how the brain got built and switched on, unless there is some special supernatural place inside the brain, it IS a machine which feels consciousness.

Now other interesting questions which are harder are: even if "we" get to the point where we can build a machine that feels consciousness, will we, in some fundamental way, understand how we did it? Machine learning seems to consist of setting in motion some software, and letting the complexity build inside the machine as it is exposed to more and more experience. If we build a conscious machine using the techniques gathered from machine learning, we may find we can build custom minds that seem as conscious as the next guy, but that no human really has any fantastic insight into what the secret sauce was that was finally found through machine learning that allowed us this capability.

OK, I'm going to go buy Domingos' book now.

Per Kurowski writes:

I absolutely believe a Universal Basic Income is needed. That could allow us citizens to solve many problems, while keeping those obnoxious redistribution profiteers away.


Comments for this podcast episode have been closed
Return to top