Datasight.io – Machine Learning for the Masses

by ncvc • 12 years ago

23 20 12 years ago

agibsonccc4• 12 years ago

It's not a bad idea, but it just doesn't seem to me that "general purpose" machine learning as a service has any real value for lay users.

I'd love to get others thoughts on this and if they would use a service like this what you would want in something like this.

I would imagine something like churn prediction as a service, face recognition as a service, sentiment analysis,... more concrete needs could be good.

The landscape is mainly APIs (wise.io,algorithms.io,..)

BigML does decision trees as a service which at least allows you to interpret the results with a good visualization.

Ersatz does neural nets as a service with a decent GUI and an API.

golergka• 12 years ago

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 that you're still in development, and I'm not asking for full documentation and a page of famous logos under "they use us", but at least a code snippet or something like that would make it easier to visualize.

Right now, I'm lost; my data is a lot more complicated and inter-connected then a mere spreadsheet, will you support it? How will it work? What pricing will be, and how will it be calculated? Yes, I understand that you haven't decided yet on all this things, but that's the problem: there's nothing to show except for the idea yet.

Anyway, I really hope that this service will work great, and I wish you all the luck with it.

swatthatfly• 12 years ago

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.

ncvc• 12 years ago

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.

minimaxir• 12 years ago

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.

agibsonccc4• 12 years ago

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.

muzakthings5• 12 years ago

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?

dave_sullivan2• 12 years ago

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...

muzakthings5• 12 years ago

Founder here. Yes of course - garbage in, garbage out.

Still, we've constantly been surprised by how well we can learn on raw CSV dumps from SQL joins or just log data.

Furthermore, small business owners or non technical folks can still understand the concept of "teaching" an algorithm: things like including demographic information increase predictive power etc.

Gonzih• 12 years ago

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.

dave_sullivan2• 12 years ago

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.

muzakthings5• 12 years ago

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.

cschmidt• 12 years ago

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.

afandian2• 12 years ago

Unfortunate phonetic naming clash. http://www.datacite.org/

muzakthings5• 12 years ago

Haha, this is great. Thanks for the link.

afandian2• 12 years ago

Seriously though. If this is a new thing you're launching you may want to consider re-naming before it becomes painful.

twog• 12 years ago

How can I get access to this? I tried to email contact@datasight.io but the address bounced

Datasight.io – Machine Learning for the Masses

Related Stories