Eight fallacies of Intelligent applications

“Algorithms of the Intelligent Web” explains eight fallacies of Intelligent applications:

“Fallacy #1: Your data is reliable

The data that you have available during development may not be representative of the data that corresponds to a production environment. Your data may contain missing values. Your data may change. Your data may not be normalized. Your data may be inappropriate for the algorithmic approach that you have in mind.

Fallacy #2: Inference happens instantaneously

You shouldn’t assume that all algorithms, on all datasets, will run within the response time limits of your application.

Fallacy #3: The size of data doesn’t matter

The size of the data matters in more than one way and you should always ask: Do I have enough data? What’s the impact to the quality of my intelligent application if I must handle 10 times more data?

Fallacy #4: Scalability of the solution isn’t an issue

You should consider scalability during the design phase of your application. In some cases, you may be able to split the data and apply your intelligent algorithm on smaller datasets in parallel. The algorithms that you select in your design may have parallel (concurrent) versions, but you should investigate this from the outset, because typically, you’ll build a lot of infrastructure and business logic around your algorithms.

Fallacy #5: Apply the same good library everywhere

Intelligent application software is like every other piece of software—it has a certain area of applicability and certain limitations. Make sure that you test thoroughly your favorite solution in new areas of application. In addition, it’s recommended that you examine every problem with a fresh perspective; a different problem may be solved more efficiently or more expediently by a different algorithm.

Fallacy #6: The computation time is known

Typically, people expect that, when we change the parameters of a problem, the problem can be solved consistently with respect to response time. A seemingly innocuous change in the data can lead to significantly different solution times; sometimes the difference can be hours instead of seconds!

Fallacy #7: Complicated models are better

Always start with the simplest model that you can think of. Then gradually try to improve your results by combining additional elements of intelligence in your solution.

Fallacy #8: There are models without bias

The choice of the models that you make and the data that you use to train your learning algorithms introduce a bias. Bias constrains our solution inside the set of things that we do know about the world (the facts) and sometimes how we came to know about it, whereas generalization attempts to capture what we don’t know (factually) but it’s reasonable to presume true given what we do know.”