# Talking to companies at ICML19

## Метаданные

- **Канал:** Yannic Kilcher
- **YouTube:** https://www.youtube.com/watch?v=hkw-WDBipgo
- **Дата:** 13.06.2019
- **Длительность:** 7:26
- **Просмотры:** 2,182
- **Источник:** https://ekstraktznaniy.ru/video/13957

## Описание

A short rant on sponsor companies at ICML and how to talk to them.

## Транскрипт

### Segment 1 (00:00 - 05:00) []

all right I quickly want to talk about kind of the interaction with corporation company reps at these conferences because to me it's still a bit of a secret or not really clear of what to do there's very different kinds of companies at these conferences so some companies I feel are there to basically show off their technology kind of wanting you to use it one example is for example graph core they're kind of new kid on the block for AI hardware in that they claim to have a chip specifically designed for the types of operations that machine learning applications do so even more specialized than a GPU and also they claim they are faster for equivalent kind of money spending than an NVIDIA GPU like a classic GPU so basically you get much more bang for the buck now for now they just offer a cloud solution I believe and they're gonna sell their cards through Dell the way it works is to have kind of a low-level compiler that will compile your model to these cards and for now you can interact with it through C++ and then tensorflow will come later something like this thing about their card is that they have an extremely large memory right next to the compute Unix this would be kind of your traditional level one cache yeah so that means that you get much faster access technically to your local variables but then they don't have any kind of RAM which means they're like entire card only has somewhat like 300 megabytes of memory but they claim they can just basically distribute if you have a large model you can distribute that over many cards and then it you'll get basically the speed up of the cards without having to sacrifice a model size another company that shows off really cool technology is a company that does lidar and I forget the name right now but when I try to look it up they so they do a lidar sensor basically that is super tiny and it costs a fraction of like a traditional lidar sensor so I think they said there's cost about twelve thousand dollars and it's really tiny and as a couple of advantages compared to traditional centuries as far as I understand their lasers are mounted on the same chip so they always point in the same direction which reduces a lot of inaccuracies so if you know I guess people who would be interested in that you know from self-driving cars and so on these are kind of the hardware demonstrations that I've seen then there's other things like there is a Wellness Center where you can get a massage which is sponsored by the big companies which is pretty nice but I'm not I'm probably too much I don't like these kinds of things too much maybe I'm just socially too awkward yeah for some companies I feel that they're just there to recruit and they don't really want to talk about what they do too much so this would be an indication of this would be a company where basically all of the reps at the booth are recruiters so not non-technical recruiters and that basically just kind of tell you what you can do as a career and not really what the company does as a whole so um I never really know what to talk about then because I don't feel like most people are interested and drawn towards interesting work and if that comes with good working conditions then that's a plus but I don't feel for many people that is the most important thing though I could be wrong and probably it's good that for some people it is

### Segment 2 (05:00 - 07:00) [5:00]

because otherwise everyone would take my jobs the ones that I like yeah so these companies will usually if there is an engineer they'll not talk about too much what they do like ours company secret and so on so the funniest one was actually the NSA - the NSA it was kind of painful because you kind of ask them so what do you do and they're like yeah machine learning like are there any you know because what I want to know as a researcher right is like is there anything I could do there that I couldn't do anywhere else right so is there any these are any unique problems that the NSA faces that actually demand new research like demand new machine learning methods or you know some kind of change ask they seem like yes there are problems like this and you ask like which problems and they're like yeah there are problems we can't tell you so everything's basically whatever I made it a game to like ask the more specific questions and watch them like oh this is classified so yeah if you're here definitely check them out it's a fun it's just fun to talk to them yeah so I feel to most companies they're really interesting I don't know more than half of them so just going up ask them what they do just get an overview over the landscape of what's neat is currently in machine learning research I think that's really useful because as an academic I tend to be very disconnected from the industry side of things and from what people actually need or want in practice so talking to all these companies is really helpful to get an overview over that yeah so but if you know a better way I know people some people are much more successful than me talking to companies at conferences I'm definitely not the best at this and if you have a better strategy let me go I'm pretty happy so far all right