Solder Voiding, Autonomous Autos, and Statistics—So Much to Learn from Dr. Ron

Estimated reading time: 20 minutes

Lasky: That would be a while. I think it could start sooner than 2045, but it's not five years away, it's decades away. I'm not suggesting that there won't be a lot of interesting things and experimental vehicles that you'll be able to buy that can do some of the things, but the car isn't too useful to us until it's fully autonomous. One of the reasons is if you have one that's partially autonomous, the car tells you, "I'm driving, you can read your book," but you still need to be looking out of the corner of your eye in case the car has a problem. A lot of people are saying they think it's actually dangerous to have a partially autonomous car, because you still need to be there and you can't completely let the car take over.

The challenge, and this is in one of my blog posts, that I think people should consider is what an autonomous car needs to do to take me from my home in Woodstock, Vermont, to Logan Airport. So I have this scenario. First, I live on a road with no lane markers. So it's got to drive up my driveway, which is about an eighth of a mile long with no road markers. The vehicle picks me up and then I tell it, "We're going to Logan Airport." Some pretty good humanized voice recognizes me and turns the car around, and we drive down into Woodstock, Vermont, but there was a thunderstorm last night and one of the bridges is out. There's a policeman now that stops the car and gives it oral instruction, “You're going to make a turn, and then a little further on the way, there's a volunteer fireman with a hand-written sign that says you have to do something.” The autonomous vehicle must be able to handle all this. I can't intervene because it doesn't have a steering wheel, brake, or a gas pedal.

We get past all the construction and we get onto Route 89, which is a major highway in New England, which has good markers and all of that. We're driving along and things are good. It's able to go through the toll booth and everything. It's early winter, we hit a brief snowstorm with a whiteout, and maybe it handles that okay. It's got to use a combination of GPS to know approximately where it is, but then it must use the road markers because GPS only gives you about 10 feet of accuracy. So it's got to use the road markers. We get near Logan Airport and there's an area under construction. Some more hand-written signs, and there are detours all over. It's got to be able to handle this, and so things that we as a human can do effortlessly are very, very difficult for computers to do. The thing of it is, we will get 95% of the way to autonomy quite quickly, but the last 5% will be very painful.

The time when the car needs to stop and roll the window down, then talk to the policeman who directs it to a place where a person is holding up a poorly handwritten sign, telling the car that there's a detour and it now has to cross one of the beautiful covered bridges we have in Vermont will be a challenge. Those are the kind of things that a human handles easily, but will be the last couple of percent of the full autonomous vehicle. That will not be here for a while.

Goldman: You don't feel it's the sudden deer jumping out? There are a lot of reaction things as well.

Lasky: Those kind of things, too. For my friends that get over-excited about what machines can do, I have one thing to say: I don't think it's so much that they overestimate what machines can do; it's that they underestimate how amazing a human being is—the things that we take for granted that are so very difficult for machines to do. So I think the full autonomous vehicle is quite a while off, and one of the reasons I wanted to share this with you today was because in yesterday's USA Today a person wrote an article saying people that drive trucks and taxi cabs need to start looking for new jobs. It's just not true, it's just not true.

Maybe their grandchildren will not be taxi cab drivers, maybe not even their children, depending on their age, but if you're driving a truck or a cab, you've got a secure job for at least 20 years. That's the main message I want to share. I felt it's appropriate because the fun thing is we are going to be the people that are going to make these electronics that must have reliability beyond what the space age needed. That's going to be the fun part for us.

For the last topic, I thought some of our readers might be interested in the polling that is going on for the presidential election, and the reason I have an interest in this is a lot of my work in electronic assembly is analyzing data. I also teach statistics at Dartmouth College, and right now, just coincidentally, we're covering polling data analysis in statistics while the polling for the presidential election is going on. You hear these polls and they'll say it's a statistical tie between Secretary Clinton and Donald Trump, and the margin of error is 3%. I think most people hear that and they say, "Well, what does this mean?” What I thought I'd do is discuss how this is done. Before I talk about that, I'm going to talk about flipping a fair coin, because it's the same statistics.

Let's say I have a quarter and—some may also find this interesting that people have actually studied this—an American quarter is a fair coin. It'll land 50% heads, 50% tails. Let's say I flip it 1000 times, so I'd expect most likely that I would get 500 heads, but I think even if you're not into statistics you're saying, "Well, you wouldn't expect to get exactly 500 heads.” If you got 496, you wouldn't think that somehow that's strange, or if I got 507, I wouldn't think somehow that's strange. Probably if I only got 200 heads, I would think that's strange. This diagram I'm showing you is something called the binomial distribution, and it shows with a fair coin essentially how likely it is to get so many heads or tails. You see, it's sort of like the bell-shaped curve, and there's a red spot here at 468, and less than 468 is red. On the high side, more than 532 is red.

What this is showing is that if we flipped a fair coin a thousand times, and we did it many, many times, most of the time we'd get about 500 heads. Any number more than 532 or less than 468 would be  statistically significant, meaning maybe the coin isn't fair because this isn't likely to happen. It's likely to happen about half of 5% of the time that it would be greater than 532 or o about half of 5% of the time that it would it be less than 468. When people on the news say it's statistically significant, they're saying 95% of the time it would be in the blue, and 5% of the time it would be in the red. They're saying it's just been decided by data analysts that if it's beyond 5% that means it's statistically significant and that it's different.

If we got greater than 532 or if we got less than 468 we'd think our coin isn't fair, but if we got something like 480, we can't say that it's not a fair coin. How does this relate to polling? We'll assume that Secretary Clinton and Donald Trump are about tied, which as I understand it is pretty close in the national polls currently. The polling people have several challenges. One is they're going to poll about 1000 people, and they've got to find 1000 people that represent the United States.

Let's say for some miracle, they do that and they represent the United States perfectly. Just like in flipping the coin, you wouldn't necessarily expect that, assuming it's a 50-50 tie throughout the country, exactly 50% would be for Clinton and 50% would be for Trump. You only have 1000 people out of the many tens of millions that are going to vote, and so you would expect that there could be some statistical variation in your sample of 1000. It's the same thing with flipping the coin—as many as 532 or as few as 468. Say we're counting the numbers that would vote for Clinton. If you got between that range 95% of the time, that's what you would get if you have a sample that represents the nation. The difference between 468 and 532 is the statistical error. So when they say it's 51% Clinton, 49% Trump, they'll say it's within the margin of error of 46.8 to 53.2%.

Goldman: Let me ask you this then, because they're polling 1000 people, but they only polled them once. How is that statistically significant at all?

Lasky: They're taking one sampling, and you could say the one sampling certainly is not as good as doing many samplings. The reason we must go with one sampling is more a reason of logistics; they just can't do more than that.

Goldman: I guess I get a little worried when they start with the “statistically significant,” because it's so easy to lie with statistics.

Lasky: You can lie with statistics, and I'm not disagreeing with you, but if you do it this way and you understand that it's done this way, and it's done by people that are trying to be fair and they explain how they did it, this is essentially the story you get. So what happens is if there's a poll that favors Hillary Clinton, her people will talk about that poll. If there's a poll that favors Donald Trump, his people will talk about that poll. If you watch the news now they’ll have an average of the polls. That makes the argument that the more people you get to survey, the more statistical precision you have, and so I like the average of polls. There are great challenges that pollsters have today; one is trying to get 1000 people that represent the country and that is not trivial. Another problem is if they call only telephone landlines, they're missing young people. In my experience, I don't know any person under 25 that owns a landline; even my adult children who are in their 30s and 40s just have cellphones.

Goldman: The other thing is that when they call, many people look at that call coming in and say, "I'm not answering that because it's a prank call, or I know it's a junk call." That's me, whether it's my cellphone or my land line; if I don't recognize the number I'm very hesitant to answer.

Lasky: One of the things when you're watching the news and they're talking about polls and they say the margin for error is plus or minus 3%, that's what it is for a sampling of 1000 people. That's typically the number that the pollsters go for—to get 1000.

Goldman: How do they know they’ve got the right 1000 people?

Lasky: That's the whole thing—getting the sample is a real challenge. That's one of the reasons why I think especially today, with all the changes in our society, it's harder to have more confidence in polls than in the past. One of the big issues is: Who will vote? Some people will argue one candidate has a better ground game and those people will go out and get their supporters to vote, or that the other candidate’s supporters are more passionate so they're more likely to vote. I'm not an expert in this, but if you ask me just as a person that thinks about data quite a bit, I think that's what's going to make the difference this election—getting the people out to vote.

Goldman: Almost in any election, but obviously when it's close it's even more important. Thank you for your insight today, Dr. Ron.

Lasky: Thank you.

• < Previous Page