Wednesday, September 19, 2018

The power struggle

The power struggle shows up in very small scale challenges, like

forcing you to use templates, projectors, printers and so on

More seriously to use powerpoint, you need a laptop or table, the right sort of cable, an electricity, which is in short or unreliable supply in many parts of the world.

noise, cost

All of this illustrates that the real problem is to do with asymmetric power, hence the power struggle really concerns the fact that the rich get richer, the poor don't.

Power struggle:

amazon, baidu, chrome, didi, electricity, facebook, google
hsbc, instagram, javascript, kdd, lambda, maps, mooc,
oed, pinterest, queen, r, s, twitter, uber whatsapp/wechat,
xkcd, youtube, zoom

comprehensive AI - or explicable ML or understandable computerised statistics

How do we know that the underlying machine that does automated decision making isn't wrong?

Well, most machine learning (ML) is about as sophisticated as stuff people used to do with
SPSS (then S, then R, and sometimes Matlab, or Python libraries) and
consists of linear regression.. occasionally something a tiny bit cleverer like naive Bayes, random forests etc etc - verifying the code, and its properties after various data have been fed in is very simple. Indeed, the vast majority of things people are doing with computers doing "AI" (not artificial intelligence, just stuff that does statistics on data, dynamically, and gets better at it, usually), are things humans did with pencil, paper, and tables, hundreds of years ago. And with mechanical calculators 100+ years ago.

People don't use deep learning in these systems (e.g. convolutional neural nets )
for much (outside of some image classification (its a cat!))... in
practice, at least as far as we can tell...there's no need, plus the
training time, data, costs (energy/processing) for neural networks
is awful and they are badly thrown by adversarial input

If you must use things like that, and are interested in their
properties, one approach (actually from the Turing inst)
now deployed in google's code, is described here:
which generalises to other systems, and has been scaled up by folks at Deepmind
to cope with systems with many dimensions and also google using a slightly different approach:

That's not verification in the sense a computer scientist would mean - its
empirical/evidential - there are some more complex approaches for
that sort of thing)....

You'd definitely need to keep checking the systems behaviour as it
is trained more as a NN can suddenly switch its behaviour if the
input is significantly novel - an llustration of this is in Wales
work in cambridge on energy landscapes - see

or you can break the system down based on a model
and generate training sets (or partition the neural net) either GANs or segmenting -

I have no idea how much anyone in the Real World uses
general model inferencers (as per MCMC/metropolitan hastings and
probablistic programming techniques) but these are relatively
transparent in the things they output too - we need to map the communities of practice. asap

This all needs documenting in a decent way because it is (increasingly) not magic/fiddle factors/pixie dust....

However, here's a reason to care about running a Butlerian Jihad:

Friday, September 14, 2018

Entangled Neural Networks & privacy preserving deep learning

There's a natural synergy between quantum computing and neural networks - a collection of entangled particles have correlated state - so instead of moving through the very large state space, they retain information with lower entropy.

so when we train a neural network made of quantum neurones (queurones), we want to increase correlation when the output vector agrees more with our goal, and decrease it when it disagrees.

so this just means generating more or less particle pairs with spin (for example), or observing one of the particles (to destroy the entanglement).

Hence we can build a very fast, high dimensionality neural net with only a few queurones, and we can also make sure that its operation cannot be observed without completely destroying its learning.

Thanks to Adria Gascon and Graham Cormode for discussion that led to this idea.