The Grand Punditess of Them All

Sharing LV’s recent talk at CogX. Of the 300-400 speakers she was, I believe, the only women (correction – handful of women) speaking on a technical subject so a huge stride forward for #WomenInStem and #IndiansInAcademia (academia in Britain especially in the higher and more complex echelons is astonishingly white). It might be shirk to say so but I suspect Vidhi might be Lakshmi in human form..

Judging from what our beloved commentariat constantly snark about me in the threads it’s astonishing she married a lightweight like me 😉

https://www.facebook.com/zachary.latif/videos/10105007111664881/?comment_id=10105012834246781&notif_id=1528957991813595&notif_t=video_comment

0 0 votes
Article Rating
21 Comments
Oldest
Newest Most Voted
Inline Feedbacks
View all comments
Razib Khan
Admin
6 years ago

Nerd

Violet
Violet
6 years ago

Aren’t you cute!?

Never thought stuff from BP would be something I would forward to my colleagues. So cool!

I wonder how it compares to cross-entropy methods in flexibility.

Violet
Violet
6 years ago

Not that you guys should discuss BP comments at home, but couple more things for future reference.

1. Inversion of large matrices is difficult but has many developments in finite element methods for large matrix inversion and storage. With domain decomposition method, this isn’t a major disadvantage for GP.

2. GP advantage of smooth twice-differentiable functions is also its disadvantage for step functions. Also co-variance can only model first order relationships between any two variables.

But the presentation appears to be for non-specialist audience in surrogate modelling, and these may not have been discussed. Sorry had to geek out somewhere.?

VRL
VRL
6 years ago
Reply to  Violet

Hi Violet,

On 1) there is a lot of work going into making GPs sparse by looking at approximations or subsets of data. I think that is more theoretically interesting than a brute force computational approach on just making the inverse part faster. (I thought the fastest algorithms for inverse were O(n^2.3728639) which is better than O(n^3) but still not much…)

2) You can create custom covariance functions to capture any shape or smoothness requirements, as long as they are valid covariance functions (psd). Also, you can have non-stationary covariance structures …

AnAn
6 years ago
Reply to  VRL

Vidhi welcome to Brown Pundits! Can you contribute articles?

Most people try to solve problems optimizing for specific equilibria, which causes suboptimal whole of system solutions because it doesn’t account for covariance. But geniuses like you can understand whole complex systems (or systems of systems for geeks) and solve for general equalibria (optimize for many variables simultaneously).

Or just write articles on how AI is changing the world.

Note I am supremely stupid as our dear beloved friend and mentor Kabir correctly said.

On estimations of subsets to make Gaussian Process sparse . . . well yeah; can you at a future point share some ways this is done?

Do people still use O(n^3) ?

You lost me (again being truly at the very bottom end of stupidity) when speaking about custom covariance functions and non-stationary covariance structures. I am badly in need of an IQ boost pill. [Go brain therapy, gene therapy and inserted bio-engineered tissue!]

Must be nice to live with someone as smart as Zachary. Especially since he tries to hard to hide his intelligence by pretending not to be intelligent so as not to make the rest of us jealous! I love and envy Zach’s deep intellectual humility. 😉

Violet
Violet
6 years ago
Reply to  VRL

@VRL,

Thanks so much for dropping by!

1. Intuitively one would think that correlation won’t be strong among a large number of x, and if there is a strong correlation you might end up with a new representative variable (which could itself be a function of all sub variables). So, sparse covariance matrix makes sense. Also FE methods use all possible methods for converting stiffness matrix to sparse matrix (by reordering etc) before inversion . Brute force is when all else fails, but makes us deliver some answers in the meantime.?
2. Yeah, non-stationarity could be interesting but then I guess it runs into same problems as non-stationary signal representation methods (specifically I am thinking of an earthquake signal in frequency domain with Fourier transform vs. Non-stationary power spectral density). There are a lot more knobs to turn for non-stationarity parameters (e.g., which extra variable should be an indicator for change in covariance structure)

Anyway, fun stuff and good luck!!
Look forward to magic ML to help people like us skip the difficulties of non-linear regression and do what we need to do. ?

Definitely look forward to more of your pubs.

sbarrkum
6 years ago

Zack you riding on coat tails and claim to fame by having a gif of you with the Punditess.

I got to the point of understanding Gaussian Process is a distribution of functions. Sampling the Gaussian process gives a function.
The rest was starting to get beyond me.

Long Island, NY upper class, Sex in the city accent. e.g. saample vs sample. Did the Punditess grow up in NY/Long Island.

sbarrkum
6 years ago

Chennai, Lalchand?
Had a classmate Lalchand, I think they were Sindhis

hoipolloi
hoipolloi
6 years ago

Punditess sounds British! Another word is Pundita.

Vikram
6 years ago

I believe Saraswati is the manifestation of Shakti you are looking for. Although, very often Saraswati and Lakshmi go together.

Nice presentation, very clear exposition.

AnAn
6 years ago

Wisdom and beauty given form. Syncretic Saraswati and Lakshmi Punditess.

You get to live with such a tejaswi (spiritually effulgent) Pundita every day? Not fair. You have way too much “privilege” 🙁

Don’t get me wrong you are a smart guy. Maybe 98 percentile on composite physical health (Sharira Siddhi), mental health (Chitta Shuddhi), intelligence (Buddhi) index of all humans. {Sorry to be the bearer of bad news, but your handsomeness brings you down a few pegs in the composite index.} Vidhi is seriously 1 basis point {top ten thousandth} material.

Sometime you need to share how you ended up with such a self assured, transcendentally wise, heavenly angel. Certainly unearned privilege based on this life-time. You must have done many truly awesome things in past lifetimes.

Vidhi deserves her own theme song. Substitute Vidhi’s name and pictures with Panchali’s in this audio video montage:
https://www.youtube.com/watch?v=Q9mVKNMcGKA

Shafiq R
6 years ago

Saw the Youtube video. Very instructive. I realize that lots of complex math is behind the general results and intuitions presented in the video but she presented them very accessibly. Your wife is a very good instructor and no doubt very gifted in the intelligence department.

AnAn
6 years ago

Violet wrote:
“Look forward to magic ML to help people like us skip the difficulties of non-linear regression and do what we need to do. ?

Definitely look forward to more of your pubs.”

Couldn’t agree more!

Brown Pundits