My First Public Tech Talk: What I’m Learning in Preparation
A while back I wrote an article titled What Standup Comedy Has Taught Me About Public Speaking and Life. That article provides a pretty comprehensive coverage of everything I know about public speaking. But even with all of that experience and knowledge, I’ve been learning a lot more in the process of preparing for my first public technology talk. Or perhaps I’ve just been revisiting, deepening, and consolidating existing realizations.
Despite the fact that I’ve been immersed in the technology world since I was a child, and having now been a professional engineer for 23 years, I somehow managed to avoid giving a public talk about technology until now. Next Wednesday, I’ll be presenting at the 2019 GPU Technology Conference (GTC) in San Jose, an artificial intelligence (AI) conference that attracted over 8,400 attendees last year. This conference happens to be organized by my employer: NVIDIA. That makes the experience a little more familiar than it would otherwise be, but also raises the stakes because I’ll be representing my company on its own turf, and I expect that many of my colleagues will be watching.
I anticipate that my talk will be held in a room that can hold at least one-hundred people. I know which room it is, and it looks pretty big on the map, but I have not yet stood in it. I’ve performed five-minute standup comedy sets to audiences of over one-hundred people before, and I’ve also presented at work to over one-hundred people for ten or fifteen minutes. But this talk will last for fifty minutes, consisting of forty minutes of presentation followed ten minutes of questions and answers.
The title of my talk is Determinism in Deep Learning. I’ll be telling the tale of my technical quest over the past year to make AI-related systems operate the same way every time. For various reasons, it’s extremely important—assuming you don’t change any parameters—that your AI “factory” produces the same AI every time. It’s equally important that any given AI will behave the same way every time it is stimulated with the same input.
The most compelling scenario is if an artificially-intelligent robot kills or injures someone (by accident of course). In that case, it must be possible to exactly reproduce the process that generated the machine and its…