\
  The most prestigious law school admissions discussion board in the world.
BackRefresh Options Favorite

Any machine learning mastermen out there??

Here's where I'm confused Machine learning models are ver...
passionate gas station national security agency
  05/13/18
in general, you can't
Laughsome crimson puppy
  05/13/18
Well then how are all these AI software companies doing it?
passionate gas station national security agency
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
Gleaned from industry data sets, I presume?
passionate gas station national security agency
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
These AI software companies must be providing very customize...
passionate gas station national security agency
  05/13/18
What about meta learning? From what I understand that automa...
passionate gas station national security agency
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
So all these AI startups are either relying on public datase...
passionate gas station national security agency
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
So if you're looking to make bank in AI, it almost seems lik...
passionate gas station national security agency
  05/13/18
Your consulting firm will struggle to find large corporate c...
hairless glittery selfie
  05/13/18
What do you mean thrash me around? Like you're saying they c...
passionate gas station national security agency
  05/13/18
Something along those lines. Essentially, it's a matter of n...
hairless glittery selfie
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
what domains are you talking about, specifically? most of...
lascivious balding dog poop doctorate
  05/13/18
There's a few in fintech that look interesting Also some ...
passionate gas station national security agency
  05/13/18
a lot of the ideas seem silly and/or will be done much bette...
lascivious balding dog poop doctorate
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
Yep. They're calling them "acquihires"
passionate gas station national security agency
  05/13/18
that's the same as being bought out, even if the motive is d...
lascivious balding dog poop doctorate
  05/13/18
Dealing with disparate data sets has always been a central i...
hairless glittery selfie
  05/13/18
Are you kidding me?! Then how come every article I read says...
passionate gas station national security agency
  05/13/18
What you're reading is a mix of marketing material and/or cl...
hairless glittery selfie
  05/13/18
Where do you think the future opportunities in AI are?
passionate gas station national security agency
  05/13/18
If you're aiming high, future opportunities would be largely...
hairless glittery selfie
  05/13/18
You really think this is all hype? Don't understand Deep...
passionate gas station national security agency
  05/13/18
It's not all hype; there are several practical applications ...
hairless glittery selfie
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
As in AGI will never be achieved or just not in the next 20-...
zippy gold new version bawdyhouse
  05/15/18
As in it's more of a philosophical question that has no real...
hairless glittery selfie
  05/15/18
interesting theory
passionate gas station national security agency
  05/29/18
...
razzle-dazzle burgundy house
  05/13/18
Interesting I was trying to think of some applications th...
passionate gas station national security agency
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
You probably don't need or want machine learning for things ...
lascivious balding dog poop doctorate
  05/15/18
Luminance and casetext are two legal startups in that space ...
passionate gas station national security agency
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
So what's the solution for dealing with disparate datasets? ...
passionate gas station national security agency
  05/13/18
This is kind of an academic question, and I'm not sure if th...
hairless glittery selfie
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
...
passionate gas station national security agency
  05/13/18
So just to sum up: AI software is fundamentally different fr...
passionate gas station national security agency
  05/13/18
You appear to be using "AI" and "machine lear...
hairless glittery selfie
  05/13/18
...
razzle-dazzle burgundy house
  05/13/18
Both are poorly defined, but AI is worse. This brings us bac...
hairless glittery selfie
  05/13/18
Look to industries where processes and products are regulate...
Contagious Drunken Pisswyrm Useless Brakes
  05/13/18
Build an AI and then ask it how to scale machine learning an...
Rose Pit
  05/13/18
...
razzle-dazzle burgundy house
  05/14/18
funny, bro
passionate gas station national security agency
  05/15/18
Actual person with ML experience here. ML is the process ...
Beady-eyed Bonkers Garrison
  05/15/18
ty
passionate gas station national security agency
  05/15/18
...
Beady-eyed Bonkers Garrison
  05/19/18
...
passionate gas station national security agency
  05/21/18
You guys might be right about AI hype https://mobile.nyti...
passionate gas station national security agency
  06/08/18
Gary Marcus has been pushing this for a while now. He thinks...
lascivious balding dog poop doctorate
  06/08/18


Poast new message in this thread



Reply Favorite

Date: May 13th, 2018 5:31 PM
Author: passionate gas station national security agency

Here's where I'm confused

Machine learning models are very domain specific - one company's data can be used to train a model, but that model won't generalize to other companies' data, etc

How does one scale software using AI then?

How could you create a software which takes a company's data, learns from it to make predictions, and then sell a software package to another company which does the same thing, just for that company's data?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042817)



Reply Favorite

Date: May 13th, 2018 5:39 PM
Author: Laughsome crimson puppy

in general, you can't

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042863)



Reply Favorite

Date: May 13th, 2018 5:53 PM
Author: passionate gas station national security agency

Well then how are all these AI software companies doing it?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042898)



Reply Favorite

Date: May 13th, 2018 6:03 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042933)



Reply Favorite

Date: May 13th, 2018 6:04 PM
Author: passionate gas station national security agency

Gleaned from industry data sets, I presume?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042940)



Reply Favorite

Date: May 13th, 2018 6:13 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042967)



Reply Favorite

Date: May 13th, 2018 6:39 PM
Author: passionate gas station national security agency

These AI software companies must be providing very customized solutions to each of their clients though I would think, in terms of what kind of data is going to be used, what the client is looking for, etc.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043067)



Reply Favorite

Date: May 13th, 2018 6:00 PM
Author: passionate gas station national security agency

What about meta learning? From what I understand that automates the design of Ml models. Couldn't this possibly be used to create something like I described?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042924)



Reply Favorite

Date: May 13th, 2018 5:53 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042899)



Reply Favorite

Date: May 13th, 2018 6:04 PM
Author: passionate gas station national security agency

So all these AI startups are either relying on public datasets or large industry wide data sets that they will then use to help customers?

Here's where I'm getting at: it seems like all the value extraction will go to large corporates who are sitting on large private datasets. How is an AI startup going to get a large piece of that pie?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042937)



Reply Favorite

Date: May 13th, 2018 6:10 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042949)



Reply Favorite

Date: May 13th, 2018 6:17 PM
Author: passionate gas station national security agency

So if you're looking to make bank in AI, it almost seems like starting up a consulting firm to help large corporates navigate this whole process would be the best route.

I've been thinking about this for a while - I just don't see how you create an AI software firm that scales like the one I was talking about in OP

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042979)



Reply Favorite

Date: May 13th, 2018 6:31 PM
Author: hairless glittery selfie

Your consulting firm will struggle to find large corporate clients unless your executives and sales team have the right connections. And even if do land one big deal as a consulting shop, the company will likely be able to thrash you around if you don't have the right backing. TSINAH is mostly right here.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043036)



Reply Favorite

Date: May 13th, 2018 6:40 PM
Author: passionate gas station national security agency

What do you mean thrash me around? Like you're saying they could demand a whole bunch of shit and then try to dick me on the bill?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043072)



Reply Favorite

Date: May 13th, 2018 6:47 PM
Author: hairless glittery selfie

Something along those lines. Essentially, it's a matter of not being as experienced in the power games played in the corporate world (and not having the resources to do so).

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043093)



Reply Favorite

Date: May 13th, 2018 7:06 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043181)



Reply Favorite

Date: May 13th, 2018 6:12 PM
Author: lascivious balding dog poop doctorate

what domains are you talking about, specifically?

most of the AI startups i have seen seem doomed to failure or being bought out by a large competitor.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042961)



Reply Favorite

Date: May 13th, 2018 6:15 PM
Author: passionate gas station national security agency

There's a few in fintech that look interesting

Also some in the sales and marketing space that look very interesting. That's the area I'm interested in

Why do you think most will go down in flames?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042972)



Reply Favorite

Date: May 13th, 2018 6:24 PM
Author: lascivious balding dog poop doctorate

a lot of the ideas seem silly and/or will be done much better by large tech companies with more resources. things like this:

https://www.offworld.ai/

even if data wasn't an issue, computational resources are and companies like Google will have a major advantage. AlphaZero was an obvious idea and could have been invented by many people if they had access to a few thousand TPUs like Google. the same goes for other research in this field.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043012)



Reply Favorite

Date: May 13th, 2018 6:16 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042976)



Reply Favorite

Date: May 13th, 2018 6:19 PM
Author: passionate gas station national security agency

Yep. They're calling them "acquihires"

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042986)



Reply Favorite

Date: May 13th, 2018 6:25 PM
Author: lascivious balding dog poop doctorate

that's the same as being bought out, even if the motive is different than with a standard acquisition.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043015)



Reply Favorite

Date: May 13th, 2018 5:39 PM
Author: hairless glittery selfie

Dealing with disparate data sets has always been a central issue in large-scale data processing. The tools around today aren't much more sophisticated than they were 20-30 years ago, but costs are lower and things are faster. Machine learning and AI are largely meaningless buzzwords.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042864)



Reply Favorite

Date: May 13th, 2018 5:54 PM
Author: passionate gas station national security agency

Are you kidding me?! Then how come every article I read says that AI is the holy grail

What's your background for saying this?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042906)



Reply Favorite

Date: May 13th, 2018 6:18 PM
Author: hairless glittery selfie

What you're reading is a mix of marketing material and/or cluelessness.

Background: I used to work as an engineer for a big-name machine learning company.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042982)



Reply Favorite

Date: May 13th, 2018 6:22 PM
Author: passionate gas station national security agency

Where do you think the future opportunities in AI are?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042998)



Reply Favorite

Date: May 13th, 2018 6:41 PM
Author: hairless glittery selfie

If you're aiming high, future opportunities would be largely built around your ability to succeed as a startup founder and ride the wave of AI hype, which is a mix of being able to raise lots of money, tons of marketing, fraud and other unethical behavior, etc.

A more realistic approach would be to solve some narrow subset of AI/ML problems that a bigger company can't figure out due to corporate bureaucracy, dysfunctional management, good engineers jumping ship, etc., and then having that company acquire your company.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043075)



Reply Favorite

Date: May 13th, 2018 7:26 PM
Author: passionate gas station national security agency

You really think this is all hype? Don't understand

Deep learning is doing some interesting stuff. Transfer learning could be the ticket to AGI

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043318)



Reply Favorite

Date: May 13th, 2018 9:31 PM
Author: hairless glittery selfie

It's not all hype; there are several practical applications of machine learning like in TSINAH's post office example. The complexity and range of data that can be processed today is much larger than in the past. In the 90s, "data mining" meant processing clicks from website users and sales of items, while today you can work with data from mobile devices, "IoT" devices, etc., and the insights you get out of it are more complex. Nonetheless, there are tons of startups out there that are built almost entirely on the hype (being in the industry, I've seen tons of this).

AGI is pure science fiction and fantasy.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043924)



Reply Favorite

Date: May 13th, 2018 10:26 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36044336)



Reply Favorite

Date: May 15th, 2018 11:03 AM
Author: zippy gold new version bawdyhouse

As in AGI will never be achieved or just not in the next 20-30 years?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36054474)



Reply Favorite

Date: May 15th, 2018 1:56 PM
Author: hairless glittery selfie

As in it's more of a philosophical question that has no real answer and will likely never have one. "AI" research today is mostly built around creating better statistical models and algorithms or simply being able to run them faster. A complicated algorithmic system may be able to provide accurate predictions and behavior in a very narrow and usually contrived set of scenarios, but it does not even begin to resemble anything like the full capacity of human intelligence, which itself we'll probably never understand or be able to adequately define due to natural limits to what we can even understand and conceive of.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36055720)



Reply Favorite

Date: May 29th, 2018 5:33 PM
Author: passionate gas station national security agency

interesting theory

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36146927)



Reply Favorite

Date: May 13th, 2018 6:43 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043080)



Reply Favorite

Date: May 13th, 2018 6:53 PM
Author: passionate gas station national security agency

Interesting

I was trying to think of some applications this technology might have in accounting (my field) but all I came to was forecasting revenue, which didn't strike me as very robust

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043123)



Reply Favorite

Date: May 13th, 2018 7:03 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043161)



Reply Favorite

Date: May 15th, 2018 2:34 PM
Author: lascivious balding dog poop doctorate

You probably don't need or want machine learning for things like revenue forecasting. Simple linear regression works better for most applications.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36055954)



Reply Favorite

Date: May 13th, 2018 6:56 PM
Author: passionate gas station national security agency

Luminance and casetext are two legal startups in that space right now actually. Both look pretty cool

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043137)



Reply Favorite

Date: May 13th, 2018 6:02 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36042929)



Reply Favorite

Date: May 13th, 2018 8:43 PM
Author: passionate gas station national security agency

So what's the solution for dealing with disparate datasets? Isn't someone trying to work on this?

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043686)



Reply Favorite

Date: May 13th, 2018 9:39 PM
Author: hairless glittery selfie

This is kind of an academic question, and I'm not sure if there's a general solution here. The closest we have to "solving" this problem is open source projects like Hadoop, Spark, etc., but these will probably be superseded by other tools eventually as costs continue to drop, technical limitations expand, etc.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043967)



Reply Favorite

Date: May 13th, 2018 10:38 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36044409)



Reply Favorite

Date: May 13th, 2018 7:58 PM
Author: passionate gas station national security agency



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043489)



Reply Favorite

Date: May 13th, 2018 8:36 PM
Author: passionate gas station national security agency

So just to sum up: AI software is fundamentally different from other types of software. I can create an accounting software and sell it to millions of users. I can scale that business.

I can't do the same thing for AI software, because machine learning is not explicitly programmed. Rather, it learns from data, which is necessarily going to vary between industries and companies. There is no one size fits all solution for it.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36043645)



Reply Favorite

Date: May 13th, 2018 9:47 PM
Author: hairless glittery selfie

You appear to be using "AI" and "machine learning" interchangeably, but they're different things.

You can build "machine learning" software that learns from data using only a spreadsheet and standard spreadsheet functions (no VBA needed).

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36044037)



Reply Favorite

Date: May 13th, 2018 10:42 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36044438)



Reply Favorite

Date: May 13th, 2018 10:45 PM
Author: hairless glittery selfie

Both are poorly defined, but AI is worse. This brings us back to my original point: they're largely meaningless buzzwords.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36044460)



Reply Favorite

Date: May 13th, 2018 10:21 PM
Author: Contagious Drunken Pisswyrm Useless Brakes

Look to industries where processes and products are regulated to be similar.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36044305)



Reply Favorite

Date: May 13th, 2018 10:32 PM
Author: Rose Pit

Build an AI and then ask it how to scale machine learning and let us know what it says

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36044370)



Reply Favorite

Date: May 14th, 2018 4:19 PM
Author: razzle-dazzle burgundy house



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36049268)



Reply Favorite

Date: May 15th, 2018 10:40 AM
Author: passionate gas station national security agency

funny, bro

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36054315)



Reply Favorite

Date: May 15th, 2018 2:30 PM
Author: Beady-eyed Bonkers Garrison

Actual person with ML experience here.

ML is the process of "training" a model using an algorithm/framework and data. This results in a model that can help you make inferences or decisions given a question (should we give this person a loan? Is this an example of fraud?).

Applying the model to solve real world problems is what gives you AI. You can build software with a model that makes decisions or provides insight given data in a standardized format. For example, take a look at the Rekognition API from Amazon for image recognition.

SO I think you're making an error in assuming that all machine learning models are highly specific to a certain problem or organization. Algorithms are very specific to training certain types of models, is probably what you're thinking of.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36055928)



Reply Favorite

Date: May 15th, 2018 5:57 PM
Author: passionate gas station national security agency

ty

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36057329)



Reply Favorite

Date: May 19th, 2018 6:57 PM
Author: Beady-eyed Bonkers Garrison



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36086006)



Reply Favorite

Date: May 21st, 2018 12:03 AM
Author: passionate gas station national security agency



(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36094333)



Reply Favorite

Date: June 8th, 2018 1:29 AM
Author: passionate gas station national security agency

You guys might be right about AI hype

https://mobile.nytimes.com/2018/05/18/opinion/artificial-intelligence-challenges.html

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36205376)



Reply Favorite

Date: June 8th, 2018 8:02 AM
Author: lascivious balding dog poop doctorate

Gary Marcus has been pushing this for a while now. He thinks the current path in machine learning is fundamentally flawed, as it focuses strictly on learning and places little to no value on innate knowledge humans use to guide decision making.

I think he is wrong for several reasons. There are pretty strict limits on the amount of data encoded in the human genome. From that base number, you have to remove everything that codes for basic cellular functions, body architecture, etc. Even genes that are strictly active in the brain will often code for things that we don't care about when engineering intelligence, such as what we find rewarding.

Even assuming what is left over is a few megabytes of data that provides priors to bias human learning, this doesn't mean we need that knowledge to replicate the functionality of our brains. It could simply be a shortcut to speed human development. The reality is that we are able to push far more data through ML systems than any human will ever be exposed to in a normal lifetime. You can have very weak priors in that situation, which is why DL systems work even though they don't resemble the brain in any substantive way.

The limitations of Google Duplex and other chatbots are not due to some inherent weakness in the learning based approach. These systems need richly detailed world models from unsupervised learning. They don't actually understand the world yet, which is why they are inflexible and require large specialized data sets to work in limited domains. I really doubt this will be a problem in a few more years.

(http://www.autoadmit.com/thread.php?thread_id=3975999&forum_id=2#36206001)