Slide 1
Slide 1 text
Machine Learning in Production
Szilárd Pafka, PhD
Chief Scientist, Epoch
LA Data Science/Machine Learning Meetup
May 2017
Slide 2
Slide 2 text
No content
Slide 3
Slide 3 text
Disclaimer:
I am not representing my employer (Epoch) in this talk
I cannot confirm nor deny if Epoch is using any of the methods, tools,
results etc. mentioned in this talk
Slide 4
Slide 4 text
No content
Slide 5
Slide 5 text
No content
Slide 6
Slide 6 text
No content
Slide 7
Slide 7 text
No content
Slide 8
Slide 8 text
No content
Slide 9
Slide 9 text
No content
Slide 10
Slide 10 text
data size [M]
training
time [s]
10x
Gradient Boosting Machines
Slide 11
Slide 11 text
...
Slide 12
Slide 12 text
No content
Slide 13
Slide 13 text
Machine Learning as a Service [...]
The bottom line on why it doesn’t work: the people that know what they’re doing just
use open source, and the people that don’t will not get anything to work, ever, even
with APIs. [...]
Amazon, Google, and Microsoft are all trying to sell a MLaaS layer as a component of
their cloud strategy. [...] The problem here is a very practical matter; the MLaaS
solutions have no customer segment -- they serve neither the competent nor the
incompetent customer segment.
The competent segment: you need machine learning people to build real production
machine learning models, because it is hard to train and debug these things properly,
and it requires a mix of understanding both theory and practice. These machine
learning people tend to just use the same open source tools that the MLaaS services
offer. So this knocks out the competent customer segment. [...]
The incompetent segment: [...]
http://www.bradfordcross.com/blog/2017/3/3/five-ai-startup-predictions-for-2017
Slide 14
Slide 14 text
- kaggle already pre-processed data (focus on ML)
- sometimes less domain knowledge (or deliberately hidden)
- focus on some measure of acc only, 0.0001 increases "relevant"
- ensembles create too complex models - never deployed in prod
- data leakage: for kaggle hard to fix even if discovered
Slide 15
Slide 15 text
No content