In this presentation, AIMS Data Scientist and Theoretical Physicist Alessandra Cagnazzo discusses what's really behind machine learning. In her session "The Optimal Life" she talks about how mathematics can be applied to optimize in business, and life in general.
The optimal life
Can math help us optimise our business and life?
AIMS Data Scientist
What do different group of people consider optimisation?
Experimental physicists (a.k.a. scientist that
have to deal with NICE(ish) DATA)
Data scientists (a.k.a. “there is 38% possibility that that is a picture of my mum,
32% possibility that is a picture of the Pope, but also 30% possibility that that is
a picture of a cheese slicer”)
It all boils down to how they choose what to minimise (o maximise),
Images from Wikipedia
Theoretical physicists study often
simplified situations, that are in their
simplicity already very hard to describe
exactly. They do most of their job even
before real data are available.
The function that the Theoretical physicist minimise
is called and it summarise how all the elements
in your theory (in this case the particle attached to the
spring) behave. For isolated, not very involved systems,
made of few elements, it is possible to write it and
minimise it, not only numerically, but alsoanalytically
(taking derivatives to find the minimum).
(in close collaboration with the theoretical physicist)
Experimental physicists can
choose among different
options that the theoretical
physicists gave them, in order
to minimise the distance
Between the data and the line.
The particles don’t start to behave weirdly because they went through a recent break-up
Start from the data, and have no theorist to help them. They study very complex situations,
where the noise is often not simply white noise.
Real data are messy, but the value of understanding them for businesses is enormous.
DATA SCIENTISTS can help you incorporating this knowledge in your business.
Minimisation of some quantity/function is still the way to go! Even a complex Neural Network bases
his functionality on minimising some quantity.
One needs to cleverly pick the function to minimise in such a way that our result will not overfit the
data or be too smooth. In both cases we would end up with miserable predictive power.
Picking a suitable function to minimise is an art
Another art is to pick the procedure of minimisation (but this is beyond the scope of this