Models in Machine Learning & Data Analytics

Photos: Models (with Fish!)

And a model (lego) of a fish:




One of the more interesting historical models I've seen is this water model of the economy in the London Science Museum (a must visit London land mark, need to go back lots of times to see everything!)  Turns out it was invented by a Kiwi. Most things are.




Here's the plumping diagram of a similar machine:



For the last 10 years I've been involved with R&D and commercialisation of a model-based tool for predictive performance engineering/modelling.  First and foremost this is model based. The model is key to how the tool works,  how data is consumed to build and parameterise the model, how the performance of large scale software systems is abstracted (what's in, what's out), how things can be understood, changed and visualised.  Also important for complexity vs accuracy tradeoffs etc.

So I've been wondering what place "Models" have in modern ML/DA and what they "look like"? Harder to find than I thought.

So far I've this page with interesting links to Visualisation of ML algorithms, not exactly what I was after but interesting.

This one is more about visualisation of results maybe? But also includes animations of algorithms (yes, animation is key to understanding many systems).  Our performance modelling and simulation tool actually started out tackling this head on, the tool allowed for changes in the model (workloads, parameters, system resources etc) as the simulation was running so you could actually play with it as if it was a working system running in real-time - time could be sped up or slowed down as well). This was one of my favourite features and key to how it was initially designed (ended up being re engineered as "batch" system which I was never happy with).

Interactive play with Tensor Flow nice.

And it seems that some people (maybe everyone) thinks ML Models are just the different ML approaches (algorithms + artefacts products - e.g. the structure and content of what is learned and what is then used to guess new results with new data).

This may be trivially true. Which models are better for understandability, visualisation of the "rules" and results?  This used to be an important question in the 1980s, black box models where you couldn't easily understand what had been learned were treated with suspicion. Now maybe there is no choice for complex deep learning models?

TODO Keep thinking