Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Datasight.io – Machine Learning for the Masses

datasight.io

Post reply on HN
Print view View original

11–20 of 27 posts

Re: Datasight.io – Machine Learning for the Masses

#11

Post by agibsonccc »

minimaxir wrote: post #5

The flaw with "machine learning for the masses" is that the data still requires preprocessing in order to derive meaningful statistically significant conclusions. That skill is as difficult and tedious as implementing a machine learning algorithm itself. More importantly, any valid ML analysis requires that the source data is good , which is impossible to guarantee with this service.

That's one of my biggest problems with "any" of the ML as a service. You can't simply claim you can make things simpler for a lay audience when there is still a lot of infrastructure required for handling data, preprocessing/cleaning etc. At that point you will need some sort of data expertise anyways.

I've found most people who need machine learning usually want control of it/hosting it in house.

Re: Datasight.io – Machine Learning for the Masses

#12

Post by dave_sullivan »

Gonzih wrote: post #8

doesn't look like solid project right now. One screenshot, few word about the idea. Nothing more. There is prediction.io which looks like much solid implementation, but I still don't like it. I don't like ML software that feels like blackbox. If I can't see learning stats/graphs - I don't trust results. There is also bigml.com, but I haven't played with this yet.

You should try http://www.ersatzlabs.com -- sign up for the app at http://api.ersatzlabs.com invite code "ersatzbeta"

It's still in beta, but it's a lot more full featured than it was a few months ago. We're trying to make deep learning easy and we provide a variety of neural network algorithms backed by GPUs. Documentation still sucks... I'm co-founder.

It's a tricky balance w/ all of our products right now (BigML, Ersatz, prediction.io, datasight.io, wise.io, more) -- how to strike the right balance between appealing to Excel power users versus people who are very comfortable in matlab or python but still don't want to build their own if there's a significantly easier option.

Re: Datasight.io – Machine Learning for the Masses

#13

Post by cschmidt »

That's very interesting. I'm working on a similar concept, http://www.predictobot.com. I'm about two week from launching my beta, so I'm not sure who is further along. If any datasight.io guys are on here, we should talk. I'm in Boston, if you ex-MIT guys are still here.

Re: Datasight.io – Machine Learning for the Masses

#14

Post by dave_sullivan »

minimaxir wrote: post #5

The flaw with "machine learning for the masses" is that the data still requires preprocessing in order to derive meaningful statistically significant conclusions. That skill is as difficult and tedious as implementing a machine learning algorithm itself. More importantly, any valid ML analysis requires that the source data is good , which is impossible to guarantee with this service.

I think value is still generated even if it doesn't solve 100% of the pipeline.

Let's say you've got 2 steps: Prep the data, build the algo + infrastructure. Let's say for the sake of argument that they're about equal in terms of difficulty. If you can drastically reduce the amount of time the second takes, that will still be a valuable speedup. And I'll bet there's tools that could be included in these types of solutions that can help with data prep also (it doesn't have to be one or the other)

Databases for instance -- there's lots of options, some proprietary, some open source -- but you'd rarely consider building your own unless as an intellectual exercise. At what point does this become true for machine learning? Saying it doesn't do 100% of my job for me doesn't mean that it won't work as a product category.

Of course, this is coming from someone with a vested interest in the category, so take that with the requisite bias...

Re: Datasight.io – Machine Learning for the Masses

#15

Post by muzakthings »

agibsonccc wrote: post #11
minimaxir wrote: post #5

The flaw with "machine learning for the masses" is that the data still requires preprocessing in order to derive meaningful statistically significant conclusions. That skill is as difficult and tedious as implementing a machine learning algorithm itself. More importantly, any valid ML analysis requires that the source data is good , which is impossible to guarantee with this service.

That's one of my biggest problems with "any" of the ML as a service. You can't simply claim you can make things simpler for a lay audience when there is still a lot of infrastructure required for handling data, preprocessing/cleaning etc. At that point you will need some sort of data expertise anyways. I've found most people who need machine learning usually want control of it/hosting it in house.

Hi there, founder here. Great point. We actually create a wrapper that allows you once finished to simply cut and paste a line from a CSV to get predictions - preprocessing, scaling, dummy encoding, PCA, etc all included.

We'll also deploy to an HTTP endpoint and you can call it and get JSON response.

Does that address your concerns?

Re: Datasight.io – Machine Learning for the Masses

#16

Post by swatthatfly »

golergka wrote: post #4

I like this idea a lot, and I am a potential customer — I'm developing an app that would heavily rely on complicated data analysis, and I planned to find a data scientist after completing first prototypes. However, there's nothing at this page right now that would convince me about this service. There's no actual examples of what your technology can achieve, there's no programming API, just a promise. I understand th…

Datacratic has been doing what you ask for a while now (disclaimer, I work for them). We do ML, audience optimization, segmentation, multivariate and nonlinear models for small clients or for some of the largest shops in the industry. But the algorithms are indeed a small problem compared to building the on-demand infrastructure, managing the dataset, and all the rest of the challenges related to managing other people's data. Trust to handle data that doesn't belong to you is earned with great difficulty, unless it comes from shady sources.

Re: Datasight.io – Machine Learning for the Masses

#17

Post by ncvc »

golergka wrote: post #4

I like this idea a lot, and I am a potential customer — I'm developing an app that would heavily rely on complicated data analysis, and I planned to find a data scientist after completing first prototypes. However, there's nothing at this page right now that would convince me about this service. There's no actual examples of what your technology can achieve, there's no programming API, just a promise. I understand th…

All good points.

The flow is simple - upload a CSV and we'll spit out serialized models, CSV of results comparing algorithms, and an HTTP endpoint where you can get predictions.

To your point about being more than a flat CSV: yes that is absolutely correct. We're starting with flat but moving towards sucking in relations (automatic joins) and constructing feature vectors from that.

Re: Datasight.io – Machine Learning for the Masses

#18

Post by muzakthings »

dave_sullivan wrote: post #12
Gonzih wrote: post #8

doesn't look like solid project right now. One screenshot, few word about the idea. Nothing more. There is prediction.io which looks like much solid implementation, but I still don't like it. I don't like ML software that feels like blackbox. If I can't see learning stats/graphs - I don't trust results. There is also bigml.com, but I haven't played with this yet.

You should try http://www.ersatzlabs.com -- sign up for the app at http://api.ersatzlabs.com invite code "ersatzbeta" It's still in beta, but it's a lot more full featured than it was a few months ago. We're trying to make deep learning easy and we provide a variety of neural network algorithms backed by GPUs. Documentation still sucks... I'm co-founder. It's a tricky balance w/ all of our products right now (BigML,…

Founder here. We also do deep learning, and we include hyperparameter optimization which is a very necessary component of training any good DBN. Feel free to contact me at will@datasight.io if you have questions.

Re: Datasight.io – Machine Learning for the Masses

#19

Post by agibsonccc »

muzakthings wrote: post #15
agibsonccc wrote: post #11

Earlier quoted context omitted.

That's one of my biggest problems with "any" of the ML as a service. You can't simply claim you can make things simpler for a lay audience when there is still a lot of infrastructure required for handling data, preprocessing/cleaning etc. At that point you will need some sort of data expertise anyways. I've found most people who need machine learning usually want control of it/hosting it in house.

Hi there, founder here. Great point. We actually create a wrapper that allows you once finished to simply cut and paste a line from a CSV to get predictions - preprocessing, scaling, dummy encoding, PCA, etc all included. We'll also deploy to an HTTP endpoint and you can call it and get JSON response. Does that address your concerns?

It's great that you're doing all of this, there still comes a time when you're making a lot of assumptions, and I believe for most use cases a good default pipeline will work great for most use cases.

I know you guys are only targeting the very simple needs of people which is great. Keep in mind I'm far from your target audience being almost a little too deep in the machine learning side. Best of luck with it!

Re: Datasight.io – Machine Learning for the Masses

#20

Post by muzakthings »

agibsonccc wrote: post #19
muzakthings wrote: post #15

Earlier quoted context omitted.

Hi there, founder here. Great point. We actually create a wrapper that allows you once finished to simply cut and paste a line from a CSV to get predictions - preprocessing, scaling, dummy encoding, PCA, etc all included. We'll also deploy to an HTTP endpoint and you can call it and get JSON response. Does that address your concerns?

It's great that you're doing all of this, there still comes a time when you're making a lot of assumptions, and I believe for most use cases a good default pipeline will work great for most use cases. I know you guys are only targeting the very simple needs of people which is great. Keep in mind I'm far from your target audience being almost a little too deep in the machine learning side. Best of luck with it!

Making preprocessors for creating bag or words vectors from free text, a special IP address feature processor, etc - these are all the ways we're starting to make the automation smarter and more of a value add to experienced folks like yourself.

What value add would you want to see?

Post reply on HN