My Journey To Understanding Machine Learning, Autopilot, & FSD
Like in many families, my kids are doing remote learning during the COVID-19 pandemic. In theory, they’re learning from the teachers during Zoom meetings and then turning their work in over the district’s e-learning platform. In practice, they often learn just enough to get lost. The result? I’ve found myself repeating the second, fourth, sixth, and eighth grades at age 36.
Like many people who took journalism and photography courses in college, I didn’t get very far with math, but I did take 4 math courses. So helping out with elementary school math is no big deal, right? But when my 4th grader needed help with long division, I was lost. When my 6th grader was doing multi-digit multiplication by hand, I ended up watching several Khan Academy videos to get a refresher.
Needless to say, the experience has been humbling.
The little tastes of humble pie I’ve been getting have led me to question my knowledge of many things I write about. When it comes to the mechanical side of cars, I’m not a certified mechanic, but I’m experienced doing everything from rebuilding air conditioning systems to frame-up restores. I know enough about wiring and computers to fix most problems and even install upgraded stereo systems. Outside of cars, I’m a certified computer technician, and I’m typing this on a computer I assembled. So far, it has been a great platform of knowledge from which to write about electric vehicles, and to some extent, ADAS systems like Autopilot, but I’m starting to see the gaps and outright holes in my knowledge of machine learning.
Given that systems like Autopilot and Waymo all use machine learning to operate, it’s important. The less I know about the topic, the more I’m having to rely on what PR teams, safety advocates, and other people with various axes to grind tell us about the topic. While I do my best to cut my way through the BS and try to factor in the biases of the sources (as well as my own biases), we are at a point where the information on ADAS and self-driving systems has gone from a trickle to a firehose, and it’s not going to let up anytime soon.
So, I’m not going to get to skip the math anymore.
Charting The Road Ahead
I’m not in a good position to go back to school, and after reading about machine learning, it’s pretty clear that building skills and knowledge is more important than getting certificates. But that still leaves the question: where do I go to get started? Fortunately, a number of machine learning experts and educators have come up with guides on the subject, like this one.
Based on the recommendations I’ve read so far, I know that I’m going to need to start learning the basic concepts and a lot more about the underlying math. To do that, I’ve started with the University of Helsinki and Reaktor’s Elements of AI. Next, I’m going to start working my way through Mathematics for Machine Learning, and then start looking at the relevant programming languages, probably starting with Python Like You Mean It. By that point, I’ll probably have a better idea of where to go next than I do right now.
Either way, I’m definitely open to the input of readers on this. Feel free to send me links to good resources in the comments!
(We will put further articles in this series in this spot, so be sure to check back occasionally, and subscribe to CleanTechnica’s feeds for updates.)
Things I’ve Picked Up So Far
As of this writing, I’ve done the first two chapters of Elements of AI. It’s not much, but I’m already starting to better understand the challenge of building systems like Autopilot and “Full Self Driving.”
Probably the biggest “lightbulb” moment so far was realizing that things we think require big intelligence, like Chess or Go, are much easier for computers than very basic human things we take for granted, like picking things up or making sense of what we see. While I watched IBM’s hardware beat world champion chess players as a kid, we still don’t have robo-butlers in all of our houses, despite the way we think of chess champions as much smarter than domestic workers. The challenges facing computer hardware and math are far different than the challenges facing human brains.
Before I really understood this, it was tempting to think of driving as “easy” for computers. It’s easy for me to drive, and I have a very good driving record, so teaching a Tesla with a bunch of GPU hardware to drive should be easy if computers were beating chess masters in the ’90s, right? Why, it’s just around the corner! Professional drivers will be out of a job within 5 years, and probably within 10 years, some jurisdictions are going to see just how safe it is and ban manual driving! Right?? Yes, I’ve seen people say all of these things (sadly, myself included a few times), and even people who should know better (or do know better) have intentionally misled us up these dead-end roads for selfish reasons.
In the exercises, I learned how complex it can be to prepare very basic games like Tic Tac Toe for automation. Just when they got me feeling confident that anything could be automated, they drop the truth bomb of the real world on me: you don’t really know enough to automate everything. The real world is a lot bigger than the nine spaces of a Tic Tac Toe game, or the much larger chess board, but that’s not the real problem. In real life, you can’t see the whole “board.” More importantly, though, the roads are always changing. Traffic, weather, construction, and other drivers are always changing the rules and the right response. And all that is before we worry about human deception that arises in road rage, fake sign hacking, anti-AI vandalism, etc.
I’m naturally an optimist, but I’ve learned that we have to temper optimism with realism when considering the potential and near-term future of products like Autopilot and Full Self Driving.
Another thing I picked up is that it’s easy to accidentally make AI-based systems look more capable than they really are with bad wording. For example, there’s no such thing as “an AI” — only systems that use varying amounts of AI and machine learning in their overall operation. Other areas, like robotics and statistics, do much of the heavy lifting in systems many readers think of as “an AI.” We are still pretty far from a real general artificial intelligence, assuming that’s even possible. We definitely don’t want to leave readers with the impression that their vehicle is alive, or even partially alive and capable of more than it really is.
Back To You
At this point, I’d like to repeat my call for reader interaction on this. If you have any ideas of resources I should be looking at on this journey toward better understanding of machine learning, please share them.
Also, if you’re planning on following along and doing things like Elements of AI, please share the insights you find about things like self-driving cars, Autopilot, etc. The more we can all learn together seeking a deeper understanding, the better off we all will be going forward!
Followup article: The Big Ways Full Self Driving & Machine Learning Differ From Our Brains.
Do you think I’ve been helpful in your understanding of Tesla, clean energy, etc.? Feel free to use my Tesla referral code to get yourself (and me) some small perks and discounts on Tesla’s cars and solar products. You can also follow me on Twitter to see my latest articles and other random things.
Have a tip for CleanTechnica? Want to advertise? Want to suggest a guest for our CleanTech Talk podcast? Contact us here.
Latest CleanTechnica.TV Video
CleanTechnica uses affiliate links. See our policy here.
