Good question.
First, a caveat. I’m not a data scientist, so my opinion is the opinion of a software developer, but of the one who has a very intimate knowledge of H2O code. The code is opensource; you can go ahead and check it out from github. You can also see that the code has hundreds of branches, and most of them have tests failing.

Why are these tests failing? Not because the changes in the branches are wrong. It happens because the whole codebase is wrong. But the company’s interesting approach to tests is statistical: if 10% of unittests are failing, it’s perceived as a statistical error.

I asked them, how do they merge the code with failing tests. The answer was - “we are optimists”. To me, it means “we are in negation”.

The code is a huge spaghetti monster; it’s not scalable, and it’s not manageable. So, if you decide to work with it, you must be an optimist, and ignore the regular failures.

Now the positive side. The UI is magnificent, written by very talented people that know how to present data. It is the strongest part of the company’s product. Also, their consultants, as far as I know, are amazing. So, if you want to rely on them helping you present your data in the most beautiful way imaginable, H2O is your friend.

Not if you want correct results from processing real data. Because they can’t even parse an avro file, generally speaking. They can’t correctly read json sent over from python. Well, you can find it all by yourself, in their code on github.

I’ve been an avid H2O user for over a year and haven’t paid a penny to use its features. H2O is fairly easy to install and run. I simply installed the “h2o” R package and managed to get a working model running within two days by following their online documentation. They also have API’s for Python, Java, Scala, etc.

Is the framework buggy? To be honest, I have indeed discovered a few bugs in the software. But within 3–4 months of reporting them, the bugs were fixed. Their tech support is usually responsive when you report bugs or post questions.

So how do they make money? Well recently they started offering a premium feature called “Driverless AI”, which packages several interesting features such as K-LIME (making machine learning models interpretable), automated hyper-parameter tuning, and GPU acceleration. Not sure how much it costs to license this.

So to conclude, H2O really is a [mostly] free and open-source framework that offers incredibly fast algorithms for machine learning. Of all the different modeling tools and frameworks I’ve encountered over the years including R, Python, SAS, Spark MLLib, XGboost, Revolution Analytics, Weka, RapidMinder and Minitab, H2O is definitely my favorite. I encourage you to try it out and judge for yourself. Good luck!

Edit: There is one potential risk that I failed to mention. H2O is a developed by a small company. If that company gets acquired or decides to change their business strategy, it’s possible that future versions would no longer be free and open-source. As a result, relying on H2O is riskier than truly free and open-source platforms such as Python, R, etc.

H2O, as well as data robot, are great if you want to solve simple common data problems and you have very little to no knowledge of machine learning nor data science and you also don’t have access to people who do. In my opinion, these platforms are very limiting and they make it very difficult to experiment with your solutions. In the case of H2O, there’s some flexibility but compared to writing your code it's like riding a bike with training wheels. If you’re afraid of falling flat on your face it's great but if you want people to take you seriously you better ditch those training wheels and learn how to ride the bike yourself.

Of course, if you have absolutely no knowledge of how to build a machine learning algorithm that won’t be a problem since you’re most likely just looking for a pre-made solution.

Another issue that I have with these platforms is that they teach people bad practices. If you look at the way they work they essentially run your data against all possible algorithms in their libraries. This gives you a fast solution but it also prevents you from learning why that solution has worked.

Overall if you want to learn how to build ML algorithms yourself or are working on a serious project I’d advise you to stay away from automl solutions altogether and invest your time into actually learning data science, statistics, math, and machine learning.

The reason I prefer H2O against sklearn is because it is very hard to integrate ML models into an existing non-Python, i.e., Java-based product. For example, to be able to use a classification model of scikit-learn in real-time by a non-Python environment, (a) you need to write a wrapper that translates an input dataset into a numpy vector where you need to handle categoricals and null values, and properly sort all input values to construct a well-formed numeric array, (b) call sklearn predict function and publish the return value as a Web service, and (c) maintain the availability of this prediction Web service in addition to the rest of your application.

With H2O, you can either use their Steam product for serving ML models, or generate and import POJO/MOJO objects (so that you don’t need to have an H2O cluster running) into your existing Java application and directly use the corresponding prediction/scoring function which supports null values, categoricals, and maps input values by name rather than by the order as sklearn requires.

My secondary reason of choosing H2O against sklearn is the memory efficiency of H2O. H2O data frames are much smaller in memory and on disk (when dumped), in comparison with pandas data frames (your mileage may vary according to your data content).

As a final note, H2O does not require (but supports for scaling purposes) Hadoop, Spark, or huge clusters to work with. It is lightweight and integrates with Python as simple as sklearn does. Of course I love and keep using sklearn in addition to H2O for experimental or complicated data science tasks. However, H2O is much easier for scaling and productionizing data mining applications.

OK, I keep the old answer but it’s basically wrong and completely outdated. H2O looks like an amazing and very impressive piece of software, it’s being developed by very clever people. On the other hand it’s still first and foremost “Big Data” system, thus you don’t need it unless you have significantly more than 100 GB of data.

——

Judging from their website, H2O is a Java ML framework built on top of Hadoop and Spark. Therefore their aim is large-scale distributed (“Big Data”) ML/analytics in enterprise setting.

So the reasons to choose H2O over Scikit-learn:

1. You develop in Java
2. You have HUGE amount of data (100s of GB at least)
3. You work for big enterprise
4. You have a Hadoop/Spark cluster at your disposal

Or something along these lines. :)

H2O and Tensorflow are not tackling the same problem.

H2O is like scikit-learn on JVM that can be run on clusters (for example using Apache Spark as backend).
It’s primary goal is scalability. If you see its API, it’s really similar to scikit’s. Also it comes with distributed pandas-like dataframes.

Tensorflow on the other hand is a library for writing computation graphs for vectorized calculations (these computations are faster if you compile these graphs, and also Tensorflow uses automatic differentiation). In this it’s more like Numpy than sklearn.

To sum up: people (mostly) use Tensorflow to implement machine learning stuff (like numpy) and H2O to actually run predefined models, and build pipelines (like scikit-learn).

Bias warning: co-founder of h2o.ai

prediction i/o is spark-in-public-cloud, via a rest api, and using mllib. I've heard discussions of prediction i/o using h2o instead of mllib, but i haven't seen anything concrete about this.

h2o is open source, cluster where-ever (public cloud, private cloud, laptop).

h2o runs atop spark, or standalone without spark. h2o runs atop yarn & hadoop, or again standalone without yarn or hadoop.

h2o can be driven via rest API, and by Scala, Java, Python and R (and all at once!)

h2o typically needs far less memory than spark for the same amount of data (but spark is working hard to rectify this).

h2o allows public-cloud cluster-on-demand via h2o Play.

Cliff

H2O.ai builds H2O, open source ultra-high performance in-memory big data Machine Learning. Resulting models can be easily integrated into standalone apps.

H2O can run on systems from individual laptops to large clusters of high performance commodity servers. Because of architectural innovations and careful parallelism it scales horizontally to very large datasets and can perform ML on them tens or hundreds of times faster than other solutions. The accuracy of the algorithms, including Deep Learning, Gradient Boosting Machine, Generalized Linear Models, Random Forest and K Means, is world class.

Its three Stanford professor advisors help keep it on the cutting edge of ML. Examples are the recent additions of Generalized Low Rank Models and Coordinate Descent solvers for GLM.

H2O can be used from Python, Scala R, a notebook style Web interface, Java and REST.

Well, let's limit AI to Deep Neural Networks, which will do for now.

The major frameworks (all are available on the web) are:

Caffe, CNTK, Theano, Torch and TensorFlow

These can be used from several languages but in general are best used from Python or C++ (CNTK will support C# and Torch is best used with Lua)

If you are starting out -> you are best off with Python as it will support all these frameworks in one way or the other (TensorFlow and Torch support Python best), and C++ can be tricky to get right.

As far as which framework, that depends to a great extent on what you are trying to do. These frameworks allow you to use pre-trained models; if you can find someone who has already trained a model with a large corpus that gets close to your intentions, you will save yourself many hours of data manipulation.

One hint before you start -> Machine learning means that you the human are the teacher. Just as your teachers worked hard teaching you; the hardest part of the work in training a machine is getting the input into the right form, properly labelled and in sufficient quantity to create a non-trivial result. Learning a computer language will be the least of your challenges.

I saw this question go by, and I thought that I’d jump in.

There is no “right” place to start. Try to start from your comfort zone because you will leave that comfort zone very quickly. It is important to start. Somewhere. You may circle at times, depending on the quality of the material you look at, and your learning style.

Find a well-annotated simple example, say text classification of some sort, say sentiment classification. There are also a number of simple image classification examples - cat/dog for instance. You can follow along with the text narration, or find a Youtube that goes through the example.

While there are all kinds of frameworks, I’d pick Keras & Tensorflow, and I’d try to start with the latest version.

Then have a look. Get the code and run the code. Observe what you can - the so called “magic” as training turns a few statements of Python into a classifier.

Then get yourself onto a newbie course. Udacity has some simple ones. There are a few free ones. Look at Jason Brownlee’s stuff. Pay for his beginner books (don’t steal them !)

Then get yourself an interesting problem to solve, and find out how others have solved it. Try to understand as much as you can.

Do as much of Andrew Ng’s course as you can.

Write some more code.

If there is time (and this won’t happen in a rush), get yourself onto a more advanced course. That might actually be affiliated with a bricks and mortar school.

Iterate as needed.

About 2 months in, some lights will go on. At 3 months you’ll be poking around Kaggle. At 4 months, you will decide that its hard but you’re hooked. At 6 months somethings are starting to make sense. You will realize that some of the postings on the web are so old as to be useless - but still interesting to see what was attempted. At 1 year, you understand the simple stuff. At 18 months, you are writing your own Tensorflow extensions, and have patience with people just starting, or people who haven’t followed a decent onboarding process, and take cheap shots at NN because then don’t understand.

Not quite the timing, but that’s how it worked for me.

Hope that helps !

JK

Customer conversations are the most intimate data a company has. I wouldn’t give mine to Google.

Outside that, pros and cons depend on the purpose. For example, we’re working with a restaurant chain on a voice strategy that integrates across general inbound, some outbound items (like reservation confirmations), and some smart speaker entertainment. Crossing ecosystems (eg Echo/Alexa/Lex vs Google) is hard.

There can be lots of benefits to conversational agents. They transcend the underlying platform. Dialogflow is a very good platform but it’s sort of like asking the pros/cons of using Angular for a specific app.

Hey Alexa…..Bring Me A Beer - Rule the Robots - Medium

Listen up, Google, et al, your pushing tensors is spreading crap. Stop.

Those chasing fancy bookkeeping of numbers just because they can? Getting themselves and others dirtier than Ma Nature would like.

Silly old guy typing here? Nope. Let’s go back and use Frenet (cohort of Hamilton) as basis for some type of symmetry study in a way that is obvious when viewed outside of the influence of money.

You know, west coast valley, chief culprit, things have been going awry for some time now. The main perturbations are from choices during the past 20 years which I had a chance to observe from my own viewpoint (tower of power of truth).

AIn’t has cause tAIn’t of a massive proportion. All done by the best and brightest. Tsk.

… I’m back with a message and a plan to model this as it ought to be done … intuition needs to be brought back to the table - not, necessarily, from altering — but, by being more respectful of Ma Nature …

Validate the idea

Like any startup, one should first validate the startup idea i.e. whether it solves a real problem.

Does it need AI

Check whether integrating AI offers better solution. Name sake AI startup doesn’t have any long term value if doesn’t solve problems with AI.

What is AI

Artificial Intelligence is a broad term encapsulating actions performed by the machines based on a mathematical model.

Creating AI involves using Machine Learning, which has been done for ~ 50 years and lately due enhancements in computing technology & algorithms its subset Deep Learning is used as well.

Machine Learning is nothing but glorified statistical learning & systems identification.

Now that basic terminology is covered, we’ll move on.

Recruiting for AI

To create something truly innovative with AI,

We need professionals who are excellent in both Mathematics and Computer Science, that generally includes PhD holders from those domains.

That will be purely research oriented, so unless the startups is well funded like Google’s Deep mind; it’s out of reach.

To apply AI in a product,

A Machine Learning Engineer with clear understanding of Match & Computer Science is required.

Tools & Infrastructure

The talent hired in the earlier section would tell what tools & infrastructure setup is required.

It generally includes Machine Learning library like TesnorFlow, PyTorch, Keras etc.

It generally includes Computers with top of the line GPUs & CPUs. In certain cases special purpose hardware like TPU.

Appropriate electricity facilities to support the computational hardware or cloud infrastructure subscription.

Data

Depending upon the type of Machine Learning applied, one might be needing large to humongous amount of data set for creating the ML model. The reason why data hoarding companies like Google, Facebook, Amazon lead in AI is not accidental.

I hope this brief answer helps any startup looking forward to enter AI space, let me know the queries if any in the comments.

More important than frameworks and tools is the process behind work you will do in your AI Startup. Frameworks and tools are dependent on what you would like to accomplish, what kind of data you have etc. The process is something more general and it helps you focus on KPI that will drive your business. That is why it's crucial to answer the following questions:

- what I would like to achieve in terms of KPI

- do I have data that will allow me to achieve my goals

- do I have a way/process to ensure data quality

- do I have a way/process to conduct experiments in a controlled way so that I deliver results in the predicted time

- what is my timeframe and roadmap for building AI / ML algorithms

As we all know development of Machine Learning and Artificial Intelligence Real Time applications is very popular now these days. Most of the people manifested their interest in learning Python programming language and developing ML and AI based Real Time Applications. According to the TIOBE Index Python is top 3rd among all programming languages. Python is High Level programming language which is super easy to learn and efficient to develop Machine Learning and Artificial Intelligence based applications.

To make this development less time consumeable and efficient Open Source Leader in Artificial Intelligence and Machine Learning ''Home - Open Source Leader in AI and ML'' launches a Python Development Framework "H2O Wave" to Develop Real Time AI Applications. Home - Open Source Leader in AI and ML announced that H2O Wave makes development of real time interactive AI Apps fast and easy for Data Scientists, Machine Learning Engineers and Software Developers. H2O Wave accelerates development with different user interface components and charts including dashboard templates, dialogs, themes and many more.

If you want to learn more about it than click on the link given in the question.