Don't rush to machine learning

red information

DISCLAIMER: This blog is fully automated, unmonitored, and does not reflect the views of Garett MacGowan. The ML model may produce content that is offensive to some readers.

blue information

This blog post was generated with a (potentially) real article title as the only prompt. A link to the original article is below.

blog header

Photo by Ilya Pavlov on Unsplash

Generated: 9/27/2021
Don't rush to machine learning, folks. That's the sound of hundreds
of VCs and angels running for the exits.


Hilarious article, by the people who can afford the time to write those words
about themselves.

> Don't rush to machine learning, folks

Hahaha, thank you for helping me realize that.

What did you think? Was the article not telling it to you like it was to
everyone else?

I thought they were trolling in that I was too dense to comprehend machine
learning haha.

I love this post, but as someone that has been going through the same struggles
for a few years now, this feels like a step backwards.

I also feel that the'solution' to this post for someone that is not
comfortable with CS is just more CS that is only good for machine learning.

I do not think that is true. I think the post shows where people are going with
ML. It is more about the tools and the frameworks and less about the details
of a specific application. This is different than the old school AI programs
that were all 'Hello Watson, Can you please do something about this'

It feels to me that the "no one is using ML" meme was largely a symptom and the
solution to that problem is not as simple as "use ML." We are in a time period
where ML makes sense, and AI and NLP-based methods are showing their utility
across a wide variety of applications, including but not limited to NLP-style
data extraction. However, these new ways aren't always as easy to use for
laypeople (and many academics still don't entirely understand the issues
involved). Machine learning has been "too hard for laypeople" for a long time,
but in practice it's only been too hard for those who tried to do it themselves
(and there still are folks out there that want to do it without help).

There have also been many tools developed for popular programming languages
that allow laypeople to perform ML.

I'm getting tired of trying to find people doing interesting work in the HN
community, and instead I keep running into people who want to do ML but don't
know it.

If I do the first step of asking myself a question (which is usually a "what
is X I haven't heard of" question); "do I know anyone who can teach me about
the subject who is doing research on machine learning;" almost every time, for
the past year, I find someone who has a research question for a different
topic. Not something completely new, but definitely something worth thinking

You did the first step:

Thanks, but I've been doing this for the past decade or so. The fact remains
I've been able to find someone working on interesting research problem on a
topic I don't know much about.

What has been really great is when someone finds a completely novel approach,
usually around neural networks and neural-nets, and someone more knowledgeable
than myself can say "ok, this is why I thought NNs wouldn't work that way, but
we can now see that they do".

Well, just asking the question should be a good start. The rest is only
suggestion. Why you got the answer? Perhaps you just stumbled upon it accidentally
or it happens so often that you get pretty good with it or it's just an
intuitive guess...

I was lucky enough to have seen a lot of these questions come into fruition.

I was looking at work with a startup that is using NNs (a pretty big one) with
sophisticated features (think of the neural nets you see in images processing,
and also, a lot of work in NLP). Basically their model is doing amazing
performance on a simple dataset. I went through with some of the more
interesting papers in their stack when the opportunity came up and someone
asked me "can you talk with this company about some research papers" and then
I was sent to research and development. I am really happy it worked out. The
guy ended up going more places than just the place I found him at though.

So yeah, yes, I understand that there aren't a lot of companies applying these
techniques, but I do think they are going to play a big part.

This seems to be a very common misconception in the ML community:

"This is a solution, and is an implementation of this solution"

The reality is that Machine learning is a tool, it's not a magic solution to
everything, that we need to use like a magic wand.

The whole point of doing machine learning is to use tools with well understood
results to accomplish a task.

If we stop thinking of machine learning as the next hot trend, and start
thinking about the practical, everyday applications, then the whole ML
community gains.


Machine learning is a tool, it's not a magic solution to everything, that we need to use like a magic wand.

This is generally true. But that doesn't help the impression of a bunch of
engineers who think that "oh, if its a tool, it must be good for that thing I
need it for and I'm gonna do it" without thinking about if it has been
validated for that tool or not. Especially when a tool that does very well
applied to a few areas of the field may not be the most efficient and effective
for others.

Not without validation, and its still a tool, just the kind of tool we have
been pushing for a couple decades.

"Oh, if its a tool, it means its perfect for all my problems."

If you are in the ML space, then you should expect that you cannot do
everything with the tool in it's current form. But you should be okay that
when you do not, you will learn the limits of the tool and move on.

Some things to add: \- there are some very nice examples in the text

\- some nice links in the text (including the book)

\- the paper (
) is online. And it is an example of doing _in the text_ things in the data.

The article gives a really bad impression about what ML is about. Even the
link provided doesn't seem to be in the public domain. Is the source at the
article linked to public domain or only to people who bought a subscription?

In other words, is anyone at HN able to comment on whether this is legitimate?

PS. I'm not advocating that the article be deleted, only that we have a better
picture of the state of ML.

It's an academic paper from 2015 (I think) which states that the text
explaining ML to 'non-ML' types can use some work.

This is a repost of the same text (that's the only link I can find), if I am
correct. It's in English (unless it's a translation/translation-in-progress)
and is available for free.


Hmmm, that is actually different, even though the link is also a dead

My problem with ML and Big Data is that they're all "big, complicated
problems" and most of them can be solved with simple statistical models. I
don't claim to be great at machine learning; my career has mostly been a
battle with Bayesian network problems for which we had a bunch of weak
statistically-based models.

Garett MacGowan

© Copyright 2023 Garett MacGowan. Design Inspiration