ldn ml_1 auto stat freeze

-#auto-stat for regr and time series models -1D time series just funcs of time -what do we need? a language of models for capturing those types of dsets -lang. of models is going to be based on Gaussian process -like GP main thing it has is likelihood proc -has variation...space {of} GPs the space can capture

That doesn't sound right check 8:56pm

-sq exp. encodes -periodic kernels encode periodic kernels -constant kernels give distrib over constant kernels -white noise kernels give you WN

-operations are addition and multiplication -multiplication gives linear x linear quadratic kernels -by closure, our language includes all polynomials

-SE x PER gives locally periodic -shape of func changes slowly over time -linear {"trend"} + periodicity <<==+>>

-analysis of dset may look like-#auto-stat for regr and time series models -1D time series just funcs of time -what do we need? a language of models for capturing those types of dsets -lang. of models is going to be based on Gaussian process -like GP main thing it has is likelihood proc -has variation...space {of} GPs the space can capture

That doesn't sound right check 8:56pm

-sq exp. encodes -periodic kernels encode periodic kernels -constant kernels give distrib over constant kernels -white noise kernels give you WN

-operations are addition and multiplication -multiplication gives linear x linear quadratic kernels -by closure, our language includes all polynomials

-SE x PER gives locally periodic -shape of func changes slowly over time -linear {"trend"} + periodicity <<==+>>

-analysis of dset may look like-#auto-stat for regr and time series models -1D time series just funcs of time -what do we need? a language of models for capturing those types of dsets -lang. of models is going to be based on Gaussian process -like GP main thing it has is likelihood proc -has variation...space {of} GPs the space can capture

That doesn't sound right check 8:56pm

-sq exp. encodes -periodic kernels encode periodic kernels -constant kernels give distrib over constant kernels -white noise kernels give you WN

-operations are addition and multiplication -multiplication gives linear x linear quadratic kernels -by closure, our language includes all polynomials

-SE x PER gives locally periodic -shape of func changes slowly over time -linear {"trend"} + periodicity <<==+>>

-analysis of dset may look like
...

-historically important data

-search starts when marginal likelihood starts going down

procession of degeneration

-student tried MCMC in space covered by these trees

-greed 'good enough'... -...for these applications

-=+cut+=-

end user:sub:endian persona -clear growing "trend" -...system tells you in words

...Systems don't "tell you" things. People tell you things. People don't give away everything by eye, I see it or I don't. That's separate for the heart of the matter

Some things are destroyed before they can be taken in hand

And the hand is fixed

And the ir.. in ... .a.`TODO:propel|purportive'processional'semi-per'cut`

-projections of the data support mixture components -"automatic statistician"

Why anthropomorphise at that point?

-"Humans get tired, we try our 5 things so I'm going to give my manager this model" -systematically can find much better kernel than you would of if stuck to Gaussian or sq exp{define:SE ICYMI}

-automatic Bayesian discovery -depending on whether you trade off accuracy or interpretability

OK Hold Up. Not necessary nope. That's a human anthropo addition the interpretability is withheld information itself is not partitioned here this is a Wager.see Pascal

-automatic ways of doing model criticism

...Well...how far are these model criticisers foregone and concluded?

-"What are the snippets that we get?" -"The model involves an ..."

Biochem

-tries to take abstract numerical things and explain them in words that wuld make sense to the reader -art why

why deprive biochemistry?

-if you have big datasets and a lot of models it's tricky because you can't run all of your models on your datasets it's too computationally expensive

-automate the allocation of computational resrources -want to be intelligent about all CPU/GPU cycles we're burning

-rational agent to trade\

NB financier's lingo

-...to trade off...

-Freeze-Thaw Bayesian optimisation