Algorithms as a Service (mikesingleton.net)

This was really interesting.. seems like a whole new take on API in a sense..

analytics, fulltext search services and recommendation services are live examples of this in the wild already.

A big problem for this would be the transfer of data to and from the API. Imagine an algo to analyze gigabytes or terabytes of data.

Also, protection of the data as it is being transfered, stored, and analyzed is an issue. This is both data integrity and also protection for privacy or confidentiality reasons.

This is where I usually draw the line at the "everything is a service" world view. It is much better to provide this as a library than as a service, particularly algorithms that run across large data sets.

However, the exception I see to this is a service that applies the algorithm across a large dataset that is owned by the service. An example of this is geocoding, where you probably don't want to store the addresses of everyone in the world in your database, but it is easy to reach out to a range of providers to get latitude longitude for your address.

This is already being done lots of places. And if you take the literal definition of the word algorithm, then all web API services are algorithms as a service. So I don't like the name, but the general idea of the post is interesting.

Thinking along these lines, something that I would find more useful would be a web service that makes a well defined managed infrastructure available to me to run my own jobs or algorithms on. For example a service I can use to submit my own map/reduce style jobs to and have it run on a big cluster of systems managed by someone else; Or a service that allows me to submit jobs to run on specialized hardware, like a cluster of systems packed with NVIDIA CUDA cards. Providers of these services could also have a library of pre-canned jobs for common tasks like text indexing, link extraction, parsing W3C logs into stats, etc. With a library like that then you've got what this post is describing and more.

I like the concept of an algorithm 'supermarket' where you get all of these algos as part of one API with various objects/methods.

But the problem is one of specialization. Surely there are other APIs out there that specialize in particular aspects of what the supermarket is providing. For example, it this AAS business is providing distance algos as well as image algos, there must be Geo Businesses and Image Businesses that can provide higher quality given their specialization.

I once tried to build a web site with content tied to locations down to the smallest towns in all countries, and I found it hard to get everything right. So I think it would be interesting to have maps, all kinds of geographical information along with some related algorithms "for lease": e.g. find the nearest city to a given location, IP-to-location, address format for a given country for building proper entry forms, phone number formats, districts/county/states for a country, etc.

The database of localities and all associated information is not only huge, it's changing literally every day and thus should be constantly kept up to date. Definitely a good candidate for the hosted service scheme.

I can see this model work for AI problems that depend more on data / trained models than actual code or algorithm itself. It also makes sense for patented algorithms (SIFT maybe), paying per instance solved might be cheaper/easier than dealing with licensing issues.

Author is under impression that implementing algorithms is hard. Real problem is analysing and understanding the problem, implementation of the standard algorithms is the easiest part. If your problem is common enough to have its implementation as service, it's very likely there is free library anyway.

Seriously, ask Dijkstra's shortest path at IOI (high school informatics olympiad), I'd bet half of students would get it 100% correct under an hour.

As what we do is effectively a limited case of this (recommendations as a service), I'll give my two cents:

  This will cause a total nerdgasm, and nobody will buy it.

There is an inverse correlation for how much abstraction is between our product and our customers' bottom-lines and how quickly they're willing to open their wallets. The proposed notion here is taking a couple steps further away from providing an end-to-end solution to a problem. Geeks love that kind of stuff. It's hard to get people to pay for it.

There are many interesting algorithms in academia, that never see the light of day. If such a service could motivate more people from academia to share software , this could become very useful.

We do this in-house in my research group already, as a way of letting anyone use fairly complex-to-setup AI systems without installing them on every machine. Even lets you use heavyweight AI from, say, Flash without doing something godawful like porting it to ActionScript, because Flash can just call out to some webservice for the algorithm it needs. Or, build a demo you can distribute to people elsewhere without having to distribute the backend as well, because the GUI phones home to get its algorithmic work done on our servers.

It's worked pretty decently, but we keep it internal because we just don't have the compute resources to offer things like SAT-solving or logic programming or planning as a service to the general public.

isn't http://www.directededge.com/ already doing "AAS"?

This would work very well for memory/computation intense algos with relatively small input and output. For example, a semantic analysis engine that accepts a URL and returns the extracted meaning from a web page. Input several bytes. Output less than a k. I think this is totally feasible and worth it as long as customers can monetize the output. To me a service like OpenCalais is AaaS. Sure they have software but the hard part of what they do is semantic analysis.

This is effectively the same proposition that web services promised us ten years ago. For the most part we still develop closely-coupled software, and I don't see that changing in most cases