Not everything that we can’t understand is AI.
The term “Artificial Intelligence” (AI) has increasingly been used to describe all sorts of systems that make decisions that are difficult to understand from the outside. This, confusingly, also embraces all of the systems that borrow their intelligence only from the fact that a smart human hard codes their behavior.
Instead, we should go back and focus on the original definition of artificially intelligent systems: they draw conclusions or extract knowledge from observations and put that internal representation to work to make new decisions. That internal representation was not created by a human but by the system itself. The human was only indirectly involved by devising the system that creates that representation in the first place.
Using this definition, there are still plenty of stories out there about what can be done with such methods: distinguishing between cats and dogs, finding objects in images, adding other people’s faces to video streams, detecting words in acoustic signals... Not many of those methods are directly applicable within real world environments, though.
What does it take to enable people to try out various different methods and settings, pick the right one for a specific problem, refine the model with new data, and deploy it as – or within – an existing application?
Visual Programming for Data Science
KNIME’s Low-Code Environment, dubbed “Visual Programming” makes this easier – rather than force users to learn how to code in the language their favorite method was published in, it allows for the code-less construction and refinement of data workflows. It enables quick exploration of alternatives and intuitive combination of tools written in different languages or facilitates running data science processes on hybrid environments.
Adding Human Intelligence
It also adds the – often ignored but unfortunately critically important – ability to ingest, join, and transform the needed data within that same no-code environment. That part already adds quite a bit of domain knowledge to the process but KNIME also allows adding manually created rules and other hand-crafted “intelligent” data processing methods, in addition to classic statistical methods and a wealth of visualization modules.
Adding Artificial Intelligence
A wide range of machine learning and AI methods is built-in but KNIME also encapsulates many of the prominent libraries out there and enables access via the same visual node-connector principle. KNIME workflows can be used to reach out to Weka, XGBoost, and H2O, to name just a few,. But KNIME also offers a series of nodes that allow us to import, design, train and/or apply a deep learning model. The default settings of most of those nodes already enable us to achieve reasonable results but in order to optimize performance, here again the user needs to understand the underlying methods – not their actual implementations.
In short, KNIME’s Visual Programming setup allows us to build pretty complex things – but at the level of abstraction a data team wants to work at. If desired, that team can reach out to the underlying programming languages, but even if they don’t want to touch code, KNIME workflows expose all of the relevant bells & whistles also within that visual framework. The data science team collaborates in the same environment and adds human intelligence whenever available and AI methods where needed. If we know more about our data, we can hand craft the data engineering side of things and leave the model learning to the AI methods. If we know less about the data the engineering part can also be automated or handed over to the AI methods to learn. And best of all, at the end the entire process can be moved into production totally automatically – from within the same visual programming environment.
AI usually doesn’t stand alone, it stands on the shoulders of a lot of humans providing the right playground and picking the right artificial intelligence. This is where the strength of KNIME workflows plays out: quickly exploring alternatives for the entire end-to-end data science life cycle.