Waiting For Self-Deriving Cars

ITProToday src

Keith Kirkpatrick is the author of an interesting CACM article on self-driving cars. It is titled “Still Waiting For Self-Driving Cars” and appears in the news section of this month’s issue.

Today we discuss why it has been so difficult to get self-driving cars started.

Kirkpatrick’s article starts:

Over the past decade, technology and automotive pundits have predicted the “imminent” arrival of fully autonomous vehicles that can drive on public roads without any active monitoring or input from a human driver. Elon Musk has predicted his company Tesla would deliver fully autonomous vehicles by the end of 2021, but he made similar predictions in 2020, 2019, and 2017. Each prediction has fallen flat, largely due to real-world safety concerns, particularly related to how self-driving cars perform in adverse conditions or situations.

An Issue

As printed in the current CACM issue, his article says the following on page 13:

A potential intermediate solution currently being tested in Germany is to utilize remote drivers to control vehicles. Vay, a Berlin-based startup, has been testing a fleet of remote-controlled electric vehicles in that city and plans to roll out a mobility service in Europe and potentially the U.S. this year. The service will allow customers to order a remote-controlled car and have it drive them to their desired destination; on arrival, they get out if the vehicle and leave it to a human teledriver miles away to either park the vehicle or steer it to the next client.

I claim this shows the issue that we have with automatic driving systems. There is a typo in the above paragraph. A typo that shows what we are up against in the attempt to make automatic driving systems.

DO YOU SEE IT?

We claim that a typo like this in a CACM article is part of the reason it is so hard to make self-driving cars. Do you agree? Or is the typo not critical? While we’re at it, does the title of this post have a typo? Read on…

Ken contributes a more difficult example of this kind, one he used while addressing the same issue of AI advances in his department’s Freshman Seminar. It relates to a later section in Kirkpatrick’s article, where he discusses the issue of

…testing to ensure vehicle navigation systems understand the complex social interactions that often occur between oncoming and adjacent drivers, or drivers and pedestrians. Generally, if a pedestrian is about to cross or is crossing a street, the driver and pedestrian will make eye contact, and will use nonverbal cues to indicate the direction and speed of their movement.

Over to Ken:

Ken’s Example

Here is the relevant portion of the White House transcript of Barack Obama’s July 2012 campaign speech in Roanoke, Virginia—the one with the notorious phrase “you didn’t build that”:

The question we could submit to Google—whose own automated-driving efforts have had difficulties with interpreting rules—is:

What does the word “that” in “you didn’t build that” refer to?

By rules of linguistic derivation, the answer seems clear:

• The word “that” refers to the most recent noun, which is “business.”

• The dashes ‘- -‘ connect two parts that refer to each other.

• The word cannot anyway refer to “roads and bridges” because that is plural whereas “that” is singular, like “business.”

• (The preceding point is not a self-contradiction because in quotes, “roads and bridges” is a phrase—singular.)

Now please take a minute-plus to listen to the actual delivery of this part of the speech (you may need to click back from 0:00:08 to 0:00:00 to get the start):

First, the transcript is missing a word: Obama actually said “that—you didn’t build that.” Perhaps having a rhetorically emphasized that right after “business” makes the semantic designation even clearer? But then notice that the part “if you’ve got a business” was delivered in a quick and parenthetical manner inside the full phrase “Somebody invested in roads and bridges … you didn’t build that.”

The interpretation that the “that” refers to infrastructure is supported by the speech’s immediate segue to the Internet, sandwiched around “somebody else made that happen.” Going out to sources, Obama was channeling a point already notoriously made by Elizabeth Warren in a 2011 speech. But without going out to sources—something we cannot expect an AI to do in the moment—there is Obama’s next sentence (also included in the speech clip):

“The point is, is that when we succeed, we succeed because of our individual initiative, but also because we do things together.”

Whether Obama intended to say that people did not build their businesses can still be argued, but that’s not the real point. My first point is that AIs based on current art for ascribing meanings not only would make that ascription, but should do so for overall consistency. And the second, larger, point at the end of my slides, where I get to hard AI problems, is this:

We base our confidence in developing AI systems on Alan Turing’s principle that whatever the human mind can resolve, a computer can be programmed to apprehend and execute. But what if important data, on fine scales, concern matters that our human minds cannot resolve?

CACM puts in a bubble Kirkpatrick’s quoting another expert that “it will take years and massive compute power to train self-driving systems to understand the nonverbal cues that pass between drivers and pedestrians.” We believe the task and need of getting cars to self-derive meanings is greater still.

Open Problems

Do these issues say anything about the difficulty of getting code right? We think so, but what do you think?

[some word changes; I seem unable to fix the endpoints of my clip on the C-SPAN website.]