0:00
and all righty well how about we uh we get started and i say hello to the internet uh
0:07
welcome to open source directions uh hosted by open teams they’re the first business to business
0:12
marketplace for open source software services um open source directions is a webinar
0:18
that brings you all the news about the future of your open source projects i’m tony fafs i’m a data scientist at
0:25
quansite and i’m a community organizer in atlanta georgia i’m super excited to be hosting because
0:31
i finally achieved my master plan i was on episode one as co-host and now
0:36
look at here i am i’m running the show david i’ve overtaken you um but i really want to welcome david my
0:42
co-host now how am i how the tables have turned hi i’m david charbonneau i’m cto at open
0:48
teams uh i’m also a senior architect at uh kwonsite and
0:53
i’m based in durham north carolina and it’s our great pleasure to introduce thomas as our guest today
1:00
hi everyone i’m thomas wiki i work on the prime c project which i’ll
1:06
be talking about a lot i’m currently involved in something related to that but i feel i want to build up a suspense to uh
1:13
towards that a little bit more so there will be the reveal for them scheming next uh so stay tuned for that
1:19
but before that i’ve been [Music] studying neuroscience at brown and then
1:26
decided that academia wasn’t the right long-term trajectory for me and
1:31
then worked at quantopian as the vp of data science there for eight years and now doing doing different things
1:39
related to privacy so thanks everyone for tuning in i’m excited to be here
1:44
well thank you for being here with us i’m super excited to talk about all of your experience and work in pymc
1:51
um but before we get into all of those details um we kind of like to start to start
1:57
these episodes off just you know getting a feel for what what what y’all are finding interesting going
2:02
on it looks like this week where uh looking at what’s your favorite most recent pr
2:08
is um which is basically all the only thing you can say in open source directions so uh thomas i was wondering do you have
2:15
a favorite pr this week uh boy do i have a favorite pr yeah should i uh
2:20
like show it to you guys oh if you’d want to that would be great yeah um let me set this up
2:28
um here we go
2:33
so unsurprisingly it’s a pull request to the prime c3 repo and it’s from
2:40
greg mingus who i’ve never heard before and he says well i’m adding this mlda
2:45
stepper and well i have no idea what that is either but apparently it’s this really cool new metropolis
2:52
hastings algorithm that they researched and then uh publishing here um well that they published and then are
2:58
now adding to this library so he’s a first-time contributor and uh so he describes the performance implications
3:06
what he did to do it the references the papers other people are excited about it and then uh there’s a lot of discussion
3:13
actually going into this but he’s just really responsive to all the feedback uh adding
3:18
documentation revising the pull request just looking at the code is also
3:24
pretty massive it’s from the alan turing institute
3:29
he’s adding it to the docs writing docs writing tests as all the nice things that we’re
3:34
looking for in the pull request and what i really also like about this is not only
3:39
the quality of the pull request which is super high the responsiveness about the author but also the fact that we are able to
3:47
attract the researchers that are writing papers on these samples and then
3:53
are basically saying well we could just put up a reference implementation but instead they’re saying well more people use it if we add it to this
4:00
library which already has all the global distributions and inference engines and now it has one more sampler and
4:08
uh so so they’re just directly adding that and that is for us the best thing that could happen right is people we don’t have to implement those
4:15
or some other people who really want to use this have to implement this but just the researcher directly goes
4:21
and implements this for us so i was really excited about that it’s about to get merged and will be
4:26
part of the upcoming release uh it’s a bit at the very briefly
4:33
it’s a sampler where you have a model that has different uh modes of approximation like you have
4:39
a model that is a very rough approximation because you may be evaluated over a very rough grid and you have a very fine-tuned model
4:45
that is more costly to estimate you can use the rough approximation
4:50
the course model and the fine-grained model together and uh basically have the
4:57
use the rough model the course model to evaluate it to find a good proposal
5:03
and then evaluate on the other one so you can just uh the cheap evaluations you’re doing on the course model
5:09
and then you’re only asking the precise model which is costly to estimate for that final acceptor reject step so
5:16
really cool and a great example of open source coming together well this seems like
5:22
what pre-print should be um and pre-publication should be i love i
5:27
love seeing this and it’s super cool that you all provide a safe space for this stuff to happen too
5:32
so thank you for sharing that um david how about you have you seen have there been any pr’s that have really
5:38
uh caught your eye this week actually yeah but primarily because i’m i’m neck deep in the project so for open
5:46
teams we’ve been building a new um kind of infrastructure so we can build up rest end points and clients for
5:54
those rest endpoints really really quickly i can share my screen as well real quick and kind of show off the project
6:00
so we call it um we call it spread and we call it scrud for semantic create
6:07
update and retrieve update and delete i’m trying to find the right window here it is
6:12
so if you looked on github and you went to open teams inc we’ve got a couple of different
6:17
projects that i can share and we’re pushing on all four of these are really really fast
6:23
so we’ve got this overall spread strategy that we’ve defined uh where what we’re trying to basically
6:28
do is make it so we can just specify json schema in some json ld context and get endpoints for
6:34
free but then go further and building on the work folks have done around json
6:39
schema form generation to make it so we can really rapidly throw together
6:45
applications but applications that are informed by the meaning of the data not just the type of the data so a great
6:51
example is if it’s a location field then maybe we want to provide you like a map selection
6:57
selector or something and so this provides a way to kind of flexibly extend
7:02
how things get rendered according to their meaning the goal being we should be able to build up things really really fast by
7:09
defining these mappings and do less coding so we’ve been building up the front end of this which includes a kind of special
7:16
caching client and an extension or wrapper around the view schema form generator
7:21
and then we’ve had another group of folks working on the back end making it so that if we start with these
7:27
schemas and these context documents we get endpoints for free um so we put
7:32
we’re going to be demoing this internally uh tomorrow and i’m really proud of the team and
7:37
really really excited super cool well hopefully you can get pro anto in there the ontology for
7:43
probability distributions probabilistic ui right
7:50
integers just randomly based on distribution but uh yeah let’s get to all of the
7:55
statistics and probability distributions that we can handle now um i’m gonna move along on my uh
8:02
on my pr part and uh i really wanted to dig into the pi mc project here
8:08
um yo yeah there’s a lot of interesting stuff there’s a lot of interesting open source
8:13
drama organizational things happening on the jupiter team compass right now i
8:19
it’s my afternoon soap opera so that’s why i like those pr’s they’re uh they’re a little juicy to me
8:26
um but nevertheless uh pi mc here um ymc
8:33
uh is a project it’s on github of course um and there’s a fifth 55 over 5000
8:40
stars on the project so it’s pretty active quite a few forks on there also and it looks like on condo forge for
8:46
example there’s over 200 000 downloads so um it’s pi mmc seems like a really
8:52
popular project and i’m excited to uh talk about it here um so david maybe you could uh lead us off
8:59
yeah absolutely uh thomas can you tell us a bit about uh who started ymc and why was it started and what need
9:07
does it fill so chris wants oh yeah
9:14
i am it was just delayed oh okay good stuff um yeah so
9:22
prime c one uh was launched by chris vonnesbeck while he was still in grad school many many
9:28
moons ago and so he uh basically just wrote this together in
9:34
pure python as just to fill his own needs he needed a sampler to run his bio biological models
9:40
and then however just because it was pure python was pretty slow so he teamed up with
9:48
david huert and they then created pime c 2.0 and that was a major improvement
9:55
because it now the likelihoods were implemented in fortran code which made it run vastly faster than the
10:01
previous version and that gained quite some popularity and that’s the level where i got involved in it and
10:07
i built some park packages on top of that and and that’s how i got
10:13
to meet chris and some other people who were involved in the project at that time like john salvation and john was really
10:21
thinking ahead back then about uh how can we push these frameworks to
10:27
really the next level and he became really interested in these gradient based samples and so that requires you not only to
10:34
evaluate the log likelihood of the model which is what time c one and two and every other framework can do you also need to
10:41
compute the derivative the gradient of that and that turned out to be really difficult so he tried doing that but it
10:47
didn’t quite work and then uh but he knew that those samples were just vastly
10:53
superior to everything that came before and he was very really the first person to realize that
10:59
these deep learning frameworks um back then they only accepted theano uh now we have tensorflow and um pyro
11:07
and all pytorch all these other ones but he was really the first person to be like well i wonder if
11:12
this tool that was built for building neural networks and can compute gradients via auditive can
11:17
be retrofitted to build computational
11:22
graphs of statistical models and turned out that was not quite possible but almost possible
11:28
so i did like the legwork to do that and then really wrote the the core of prime c3
11:34
uh based on theano and just like now this is obvious right like we have
11:40
tensorflow probability we have all these other packages that are building on these graph computational graph libraries but back
11:45
then no one had even thought about it so for me there was really just like incredible foresight on on his part
11:51
to do that and then also to like really write the core of time c3 a lot of that still exists today and i remember
11:58
talking to him a lot about this and being extremely excited because i was like this is the coolest thing ever
12:03
and i contributed a few small pieces and um and then
12:11
so he built this prototype but then sort of moved on to to queen of pastures
12:17
but um i was just uh so excited about it that with correspondence back basically we pushed that over the finish line and
12:24
turned this into like a proper 3.0 release that people could use and back then it was still pretty small
12:31
but so slowly we started getting more contributors and more users and just very organically
12:38
grew over the last couple of years to now we have or over 20 core contributors and
12:43
we named some stats of download so we have a lot of users in academia there’s over 700 citations
12:49
to our original publication and it’s being used at um all the big and small companies so we’re
12:56
every day we’re learning about new companies that are using pine23 so that is really motivating and inspiring to us
13:02
and really the way the reason why we develop this is well we all just believe that basis
13:09
statistics is uh the the best way of doing statistics that’s what i would say
13:14
to me it’s the first time that statistics uh which i have many classes on but for me that was the first time it
13:21
really actually made sense we’re like oh i don’t have to like derive all these different estimators for the student t-test and whatnot i just
13:28
can write down my model and then i like hit the inference button and it works and i can change the model
13:33
like and it’s all laid out there in code like this is incredible and so for me and all other data
13:39
scientists it’s just really empowering to be able to build my own statistical models and
13:47
get uncertainty quantification and all these very nice things be able to include my expert information or get that from the
13:53
stakeholders who are interested in the model and have their own opinions on what sort of should go on so i can incorporate
13:59
that into the model too and and solve very targeted problems on very
14:07
richly structured data set problems so the contrast that with machine learning for example where
14:13
well the problems that you’re solving there is you have a table of rows of the data points and
14:20
then you have a bunch of features and then you have labels or it’s a regression type of problem and
14:25
you just if that is your problem right and you only care about prediction amazing just go ahead train xg boost or
14:33
psychic learn like it’s amazing but if your problem falls outside of that domain becomes really difficult
14:38
right um you have hierarchical nested structure in your data or it’s time serious like all these kind
14:44
of things are just really common problems right and uh they’re not often not very well suited by machine
14:49
learning which is taking up so much attention um but i think more and more these methods are being
14:55
realized it’s just very elegant powerful methods to solve these problems and
15:01
yeah they with these new samples we can finally actually do that so that was great i love all the history
15:08
um and it’s deep but let’s let’s ease into it a little bit i’m just curious about the name clearly
15:15
all of that history and the name means something can you tell us a little bit of the history of the name and kind of
15:20
the logo uh just kind of give us some you know branding context around uh the
15:25
pi mc project um so prime c just literally stands for python
15:31
and mc for markov chain monte carlo uh i think at some point
15:36
chris debate called it pi mcmc but there was one bit long in the tooth so time c it was and is
15:44
and i don’t know who first like brought in the rocket ship of the logo
15:49
um but when we uh did our logo for prime c3 on 99 designs we just uploaded the
15:56
previous one and every designer just really latched onto the rocket and i mean i think the rock is kind of cool
16:02
right that’s the spaceship uh it has been used at spacex actually to build rocket ships so that’s fitting um it’s uh yeah it takes takes you to
16:09
the moon it looks like a distribution i never made that connection but that’s
16:16
excellent yeah like a normal solution almost perfect well um are there any alternative
16:21
projects out there that you’re aware of no no i’m just kidding um of course
16:26
uh so the other really big popular uh an amazing project is stan um they are uh it’s written by andrew
16:35
gelman and a lot of other really smart people has come up about the same time as prime c3 actually
16:40
and they also use these gradient-based samples which is the major revolution and we really adapted to
16:46
these guys they have been really helpful and a lot of features in pi mc3 are directly
16:51
inspired by uh what the stand guys have been doing um the reason why i really like prime c is
16:58
just you can write your models directly in python and for me that’s just more natural way i don’t have to like
17:04
enter the data here and get results out there and there’s this weird wrapper thing going on of course the downside is you can only
17:09
use it from within python so if you’re in our user you can’t really use prime c3 but you can’t use that
17:16
and then since then there have been a whole bunch of other really cool interesting packages uh that i i would say are more
17:23
experimental or targeted at researchers just because they’re a more lower level and don’t really have
17:30
all the nice ux uh that i would say pimc has and those include pyro
17:36
um or tensorflow probability um some really cool uh work is being done there and they’re really pushing the boundaries and
17:43
pushing it also in the respect to how large scales these models can be
17:51
okay cool it’s people hear about all those different things oh tony well i’m muted accidentally okay
17:58
so thanks for thanks for that description uh is there something uh can you tell us a little bit about the technology that by mc’s
18:06
is built on and how it kind of differentiates itself from those other tools
18:11
right so prime c is built uh times a3 is built on theano which was the very first deep learning
18:17
library that allowed you to write down any neural network and then it would
18:25
take that pour that into a computational graph from the graph structure you can do all
18:30
kinds of simplifications you can compute gradients and then you can also compile that down to c
18:36
or to the gpu and then make it run really fast for pioneer c3 we’re really benefiting
18:42
from that because the whole code base is just python right imc2 had these fortran wrappers
18:47
that made installation really really difficult now it’s a pure python library
18:53
and it’s fast because well we have this compiler in the theano library that is doing that
19:00
stan has wrote that mostly themselves it’s written the course written in c plus plus so
19:07
they’re already basically been taking that and compiling that themselves so we don’t have that burden and
19:14
and we’re getting like gpu support for free which is kind of nice uh since then like all these other
19:20
packages have really followed in those footsteps of using computational back-ends now thiano
19:26
is pretty old by now i still think it’s one of the best actually but nonetheless uh people are not as excited
19:32
about it as they were about tensorflow or high torch or now jax um yeah um so that’s
19:40
what i would say is that’s the the technology that we using
19:45
who’s who are the current maintainers of the project the day to day keeping it going well really it’s
19:52
it’s a lot of people um so uh i’m uh i i don’t want to list them all just
19:58
because i would forget some um and so as i said there’s over 20 people who are like active on it
20:05
so if you just browse github or you look at the commit history that would give it much more justice but
20:10
yeah there’s a lot of people that have been involved for a really long time and they’re all extremely talented and are really
20:16
pushing the project forward so um one thing that i’d like to correct is a lot of people think that i’m
20:22
either the author or like the person who like has done the most work neither of these are true um like it’s
20:27
it’s an effort by a lot of people and and the main author of it is john salvatier who like
20:33
wrote the initial prototype so credit should go to everyone there it sounds like it’s a real vibrant and
20:39
active community and we love to see that that’s exciting when we get to talk about that well so
20:46
clearly statistics apply to a lot of people or at least i hope they realize that uh um what communities do your users and
20:54
your contributors come from or is that another thing where you can’t list it because you did you’d skip some people
21:00
too that’s exactly right it’s just too many so one thing that i love to do
21:06
is just browse the papers that are citing pioneer c3 so because like all of those are papers
21:13
that use it for them um for their science so a lot of it is
21:18
astrophysics uh biology seismology zoology um
21:24
like just insane like chemistry it’s really from like domains i haven’t even heard about
21:29
um so there’s a toolbox uh so it’s being used in a lot of papers
21:35
directly but there’s also two boxes built on pine c3 uh one of those is for example actual planet which is a toolbox used to
21:41
discover stars like planets outside of our solar system just like that’s totally wild i mean it’s
21:47
incredible to have that amount of impact right and then like and use all these different domains um there’s beat which is like for
21:53
earthquake detection uh and and just many more so yeah it’s really exciting and it’s used
22:00
in yeah i haven’t found many domains where it’s not being used um so that would be easier to list and
22:06
uh and also just in companies so a lot of companies are using it surprisingly a lot of them are using it
22:11
for simple things like ad tests um so that seems to be very popular reason but also there um in finance so i
22:18
used it a lot at quantopian to do portfolio allocation so that is really just a testament to the
22:24
personality of well statistics in general is you can just use it for whatever data science problem
22:31
you have i think for a lot of those if it’s more structured as i mentioned before
22:37
it could be a great fit
22:42
thanks so much yeah um let’s see uh is the project participating in any
22:48
diversity and inclusion efforts today um so yeah we definitely are very
22:53
cognizant of that and are um talking internally a lot about that and a lot of our members are active in
23:00
those communities so um ravine for example is um
23:05
is being very active and like presenting that at the pi ladies la meetups and those type of
23:11
things so yeah we we love those initiatives and uh where we can we support them
23:16
that’s fantastic oh you have mute again tony i know i
23:23
keep playing i got a fan on here man yeah it’s a little i apologize uh
23:28
all right that’s right um yeah so thank you so much for that introduction i think now it’s time to get to the goods
23:34
though um maybe we could shift over to a uh project demo so thomas if you kind of
23:41
get your screen set up and uh walk us through some things um but in
23:46
the meantime i really want to thank our sponsor quansite the good folks
23:51
good folks that keep me working in science and day to day on aquansight we’re creating value from
23:57
data um and uh this is brought to you by kwanzai uh we’ve got a lot of open source stuff
24:02
going on oh yeah one thing uh that i also in
24:09
terms of like which domains are using it uh there’s this website called rt.live
24:14
which is a very timely website for estimating the reproduction factor of covet 19.
24:21
so here for every state in the s you get an estimate of the of the reproduction factor and of
24:28
course for basins we have error bars or uncertainty intervals around this and you also get a timely estimate for
24:34
this um so the model was built by
24:39
kevin systrom and thomas vladick so kevin systrom is the co-founder and former ceo of instagram
24:46
so that was really interesting uh sort of thing that happened where he was just like oh yeah let me just build
24:51
this model here i’m not really an expert but i’m going to use prime c3 to uh to to do that and we helped with
24:58
some technical support and so the model is also open source so you can check it out
25:04
and really look at the time c3 code that does that and uh but yeah so
25:11
the model itself is actually really sophisticated and interesting there’s also a tutorial
25:17
here and so that is i think
25:25
i would argue one of the best models out there currently for estimating the rep production factor if you want to show
25:30
this network you might have to jump to envy fewer because github yeah you know github’s
25:36
just like sometimes right what is up with that um okay so
25:42
um prime c3 let’s just like take a brief look um and tell me if i’m going too
25:49
slow and you want me to speed up i’m gonna just try and do a high-level demo to
25:57
that’s good everybody the pace of life is far too fast for how little we leave the house i could i couldn’t agree yeah um
26:04
so uh yeah here is the website and um colin carroll did an amazing job with
26:11
that so he really uh just took our previous website which was sort of a little bit bold and really revamped
26:17
that and um so yeah you can see barely this is
26:23
probably the simplest model well not the simplest but a simple model that is just a linear regression and
26:29
let’s say we have some data x um some of these are the
26:35
coefficients and then i have some y’s that i want to model that i want to explain with my linear model i set up
26:42
a model context so this is the spoiler plate code and will be for bookkeeping
26:48
every model will look like that and then we define our parameters so here we have the
26:54
weights uh of my linear regression and i have a noise parameter that’s the variance
27:00
or the error term and here um because we are basins
27:07
we don’t just define the parameters we also define priors place on them and
27:14
here for the weights i’m now choosing a normal prior so pi mc3 comes pre-equipped
27:19
with a lot of the probability divisions that you want to use so this prior basically says without
27:25
having seen any data what do i believe a likely
27:30
coefficient for this linear regression to b right so i haven’t seen any data but i
27:36
still might have beliefs about that so here my belief is that the coefficient is going to be centered
27:42
around zero with a standard deviation of one and then for the noise well i have some
27:49
prior information right i know that the standard deviation of my normal likelihood can only be positive
27:55
so that’s why i’m choosing a sort of gamma distribution i mean a name just before i did here it has two
28:02
parameters and then we’ll just specify the shape of that gamma distribution and now i’m building this model together
28:10
and say okay i assume that my data is normally distributed again i give it a name and then my mu
28:18
the uh the mean of that data is now going to be a vector
28:24
which is the output of my data x and the dot product with the weights so
28:29
basically this is yeah just follow the math of the linear regression and then my sigma is the
28:36
noise parameter here with the observed keyword argument i give it the output data that’s supposed to model
28:42
and then this concludes the model specification and what’s happening behind the scenes
28:49
here is that we’re building up this theano compute graph of log probability so nothing’s actually happening but now once we have that
28:56
model specified we can do all kinds of things with it and inquire into it one thing is we might just want to
29:02
sample from the prior like what does the model think is um
29:08
what the prior what kind of patterns does the model think are
29:13
plausible so this is the prior predictive then i can also do that is mostly what
29:19
we do is we sample from the posterior which is basically what are my what is my belief
29:24
in my parameters after having seen the data so this is usually if you run maximum likelihood
29:30
estimate right you’re just fitting the parameters to the model now here of course we’re bayesians so we will have
29:36
not just single numbers as a result but we have probability distributions and i will show what that looks like
29:43
and then we can also sample from the stereo predictive set so this is just basically the very core
29:48
intro of how you can do this model but really the power of the approach comes by being
29:55
able to also build much more complex models so if you click on examples
30:00
you see that there’s all kinds of different ones so there’s various ones on linear regression models
30:08
and hierarchical regression which is a very powerful tool um and also more advanced ones
30:15
um there’s different case studies that you like for example this is a very classic hypothesis test that you might
30:20
want to run and also more advanced one like goes in processes which are very
30:26
flexible regression type models non-parametric ones there’s ordinary differential equations
30:33
examples mixture models so everything every type of model you might want to build we have an example of how to do that so
30:39
this is really a treasure trove of learning about this and what i also wanted to show briefly was
30:49
an example model um about this uh yeah i mean i looked at the docs last night
30:55
they were gorgeous um and i was like oh i learned some things i was like oh what’s this distribution
31:00
and it was fun to like you know kind of poke around there um how does this help you
31:05
teaching statistics it seems like you can teach statistics in a much more like rich and visual way
31:11
than perhaps like the old way uh it’s funny to say that um so one
31:17
thing that i’m doing is
31:22
building a course that does exactly what you just said um so i think there is a much
31:28
better way to teach statistics um by using intuition and
31:34
little graphics first right so i think a lot of statistics books and resources are just extremely math
31:40
heavy right so for someone like me coming from software engineering it was um it took quite a while really to um
31:48
like go through all the math and really understand things at an intuitive level once i did i noticed that like oh well
31:54
this is actually not a complicated concept at all like a lot of statistics actually really simple once you sort of have an
32:00
intuitive understanding for that and with this course which i’m currently creating
32:06
i’m basically want to teach statistics in this way so if you’re interested in
32:11
that go to tiwiki.org and and enter your email address and
32:17
then you’ll be updated on my progress on the course so yeah i think there’s a better way to teach that
32:23
i’m currently working on on providing that super cool thank you
32:32
um so one thing one idea that i just wanted to plant which for me was really
32:38
when i first grocked the power of basin statistics was this model where we are looking at
32:44
the returns of the s p 500 over time so these are just daily returns and one thing that you can clearly see
32:51
is that volatility right how wide these swings
32:56
are in both directions really clusters over time so here in 2009 right during the financial crash
33:02
there was a lot of volatility in the market then it started decayed and well we don’t plot it until today but during
33:07
covert 19 definitely a lot of volatility as well so how would we model this well
33:15
returns we might assume are normally stupid but that is actually a pretty terrible
33:20
model just because we’re having these wide swings so you can use a student t distribution
33:26
which has more massive details and accounts for these uh sort of black swan events or just like
33:32
much more uh these tail events but also what becomes clear is that if
33:38
we just fit a student t distribution right or we can just think of our normal distribution
33:44
that the standard deviation of that distribution will if we just fix that over time it’ll be a
33:51
terrible model so instead what we can do is we can allow that to vary using what is called a gaussian random
33:57
walk so that is a specific type of prior that we can place and that will allow the standard deviation to change ever so
34:04
slightly over time and this is what that mode looks like so we have this custom random
34:09
walk distribution in here and then we just feed that into the student t distribution and with the power of base and
34:17
statistics we can just fit that you just call the sample function we don’t have to do anything else we look at the posterior and then we get
34:24
the volatility over time and how it’s changing and what i just want to highlight about this
34:30
is this model is actually crazy complex so instead of just having one standard deviation parameter and one mean that
34:37
we’re estimating over this data set now we have a standard deviation parameter for every single data point
34:43
right so we have more parameters than data points here but it still works because of the
34:49
constraints we’re placing in here so because we’re assuming that this has a gaussian random walk and that it can’t just change
34:54
really nearly at every data every time point but will be constrained to a previous point uh we can still fit that
35:00
model so this is really now yeah as i said quite a complicated model but with the nut sampler this works
35:07
really well and we can just estimate it easily
35:14
so that idea we can just extend to other ones like let’s say we have a linear regression that we would do but over time so these is the this is
35:21
the price of two stocks over time um and you can see that uh they are correlated right but they’re
35:27
sort of changing over time we can just apply that same trick um so
35:32
this is if you just fit a linear regression to it but we can apply that same trick of the gaussian random walk
35:38
and say okay my intercepts and my slopes are allowed to change over time according to this random walk so they’re
35:44
going to be tied and distributed according to the intercept at the previous time point
35:50
and again we’re using that crossing random distribution i’m well aware that like this is not enough to rock but i just want to show
35:57
the high level idea of this and then again we just call the sample function and we see that the intercept is
36:03
changing over time the slope is changing over time and now with just a very simple change to our original regression model
36:10
now we get a regression that is also allowed to change over time so this is basically now the
36:15
new model and so that idea can just easily extend to all kinds of topics um like here i have a blog post
36:24
also on my blog for doing this with a linear sorry logistic regression
36:32
so if you have data that is also changing over time you can build a logistic regression model that is also
36:37
using the gaussian random walk and then fits a classifier where
36:42
intercept is changing over time so that’s pretty cool and then you can
36:48
also extend that to a neural network model your data looks like this so now we can’t linearly separate that
36:55
right so uh this is like the half moons data set and changing over time you can build a neural network model where the weights
37:02
are allowed to change over time and then also they will build a classifier a
37:07
non-linear classifier that is able to separate this non-stationary part of us
37:13
so those are just some high level points that’s gorgeous
37:19
david wasn’t i right when i said that they have great documentation i know i i know they already knew that but
37:24
aren’t there aren’t these docs beautiful with the videos and the pictures and equations and stuff like you’re learning
37:30
from them and you don’t understand the equations you get a feel for what’s going on in the math just because of the
37:35
visualizations which i think that that’s critical yeah yeah so uh maybe we shift gears and start moving
37:42
over into the roadmap discussion and use this as a jumping off point um i really want to talk about y’all’s
37:47
documentation uh it seems like documentation’s like a good measure of the health of a community and
37:53
i would say your community looks very healthy by your documentation can you tell us a little bit about that
37:59
and how you’re able to interact with the community and how you use this to kind of build
38:04
community perhaps certainly so documentation has always
38:09
been very important to us and well i think everyone understands how important it is but nonetheless it’s often not the most
38:16
exciting work one thing that i think we benefited from
38:22
is that people love to write blog posts about the cool models that they built so a lot of these example notebooks that
38:30
we have come from uh just um notebooks and blog posts that people
38:35
have done and then when i find a cool one i often would reach out to the author and be like hey this looks really great
38:41
uh do you want to add this to the documentation
38:46
so this is how a lot of this was built um by now we when people submit pull requests and
38:52
there’s a lot of people submitting these pull requests from all kinds of different places
38:59
we have a requirement that for example if you say okay well open a new pull request um
39:07
oh well whatever there is a checklist of what you need to do and it includes documentation so every
39:15
new feature that gets included has to have an example and has to have documentation so um just by placing status on that
39:23
it’s very important um but we’re also lucky that because we have quite a few
39:28
contributors from different people so here i just noticed this oriole uh abril who’s a core contributor he
39:35
does a lot of work on this and is doing fixes on this so we have quite a few people that enjoy
39:42
doing documentation and is very important and it’s a nice way to get visibility if you um do something so it’s very easy for
39:49
people to get started on and then transition to more deeper parts of the code base
39:56
oh and the last thing that i want to mention then is there’s also all kinds of other
40:02
materials so you go to books and videos [Music] we have a discourse forum where people
40:07
can ask questions uh we have a conference so we have
40:12
time c con which will happen later this year which i hope everyone will submit to
40:18
talk to and also and also attend
40:26
so i’m really excited about that i’m just posting the link uh so yeah there uh well it’s really
40:33
about community that’s uh for us really important thing and i think part of why we have such a healthy community and
40:41
here chris our you know benevolent dictator for life is going to give the keynote and we have two other ones which are really
40:46
excited about which we’re going to announce soon and um but yeah and then uh what i
40:52
actually wanted to show was uh that that there’s also a lot of work going on
40:57
into porting either porting books of people who have
41:03
um um already written books so there’s bayesian methods for hackers which was
41:08
written for prime c2 and reported to prime c3 statistically thinking which is an excellent book uh which was written for
41:14
stan but reported it to prime c3 uh oswaldo wrote book for prime c3
41:20
directly targeted on time c3 there’s all these resources and then there is uh all kinds of talks
41:27
on youtube that you can watch so yeah we’re just uh
41:32
we’re very active in this domain and i think it is a big um reason for the health of the
41:39
community and the project
41:44
this is fantastic it’s kind of inspirational actually just in terms of uh looking at how you’ve structured this
41:52
and the the material you produced it it gives you an idea of how to
41:57
if you wanted to build your own community and foster it this is a great resource um you know in addition to just
42:02
being a wonderful resource for the pi mc3 community absolutely um well
42:09
super cool uh in there maybe maybe we uh we move along and we saw cool stuff
42:15
y’all have where are you all going uh now so maybe you could tell us a little bit some
42:20
future directions some places where you might are looking for some more contributors because pi mc seems like a healthy place to be
42:28
contributing yeah definitely so if you want to get started um we are excited to
42:34
to have you um one good place to start is to just look at the issues we have different labels um so for
42:41
example with a beginner friendly label and you can just look at what is there this is like low hanging fruit that’s
42:46
really helpful for the project but uh is easy to do and but also always very eager to to
42:53
help people one trick that i started doing which is more uh more successful than i thought it
43:00
might be a lot of people are writing issues where they’re saying oh this doesn’t work and i’m like oh okay um
43:06
like the problem might be here and there do you want to do a pull request to fix that and then like oh yeah sure
43:12
let me give that a try so a lot of people were getting that way but just basically being lazy ourselves
43:17
and asking people to to chip in and um i would say like 80 to 90 of the people
43:24
percent of the cases people are really excited to do so so that’s a great way um
43:29
the um the direction that i’m most excited about is one of them is well there’s prime c4
43:37
which um i want to talk to next um but before that uh there’s all kinds
43:43
of different things going into prime c3 as well so one thing uh before i notice that there
43:49
is that we building on thiano right thiano as most people were probably aware of
43:55
uh so the the original authors stopped developing it uh which is a bummer uh because it
44:02
really was a great package and sort of was the back end that we built prime c3 on now
44:08
um this was one of the reasons why we started prime c4 but we also just realized that the existing com user base
44:15
we have is just huge right and and empire c3 is really already such a
44:22
mature package and does really almost everything we would want it to do so what we’ve done is we forked theano
44:29
so we are now maintaining ghana and pushing it forward ourselves so it’s not dead and one of the directions that we’re
44:35
pushing it forward on which i’m incredibly excited about is uh to make
44:41
it use the jack’s back end so jax is a new framework similar um to say tensorflow but much
44:48
more lightweight and better in many aspects i would argue and we can make thiano use jax as the back
44:55
end and that way we get access to all the modern technology there and
45:02
to things like gpu support or tpu support and we don’t have to worry about the um
45:08
the c code that is in the theano library instead we can just use jaxx which does all these amazing compile
45:14
time optimizations and everything but still really on piano so that way um we’re actually solving all the problems
45:20
we still have theano which is even today i think uh on par with all the other
45:25
projects and now we also are on modern infrastructure basically and don’t have to rely on the theano way of doing
45:33
things that was amazing like 10 years ago but now is a little bit
45:38
old a couple of other packages that i’m
45:44
really excited about one is bambi uh oswaldo martinez wrote the book is maintaining that
45:49
um and it’s basically allows you to specify all kinds of uh linear regression models
45:56
as you would in r so just with a single string so it’s very fast and you can use hierarchical models
46:03
another direction that is really promising is adrian zabalt wrote sun ode which is a
46:10
wrapper for a ordinary differential equation library and these ordinary differential
46:16
equation models happen a lot in epidemiology so they’re active today about
46:23
so these srr models all oes and so this library just makes it extremely fast and it’s
46:29
uh integrated with pi z3 so that is another great one um that’s cool hey of course
46:36
i got to give you uh i got to give you a choice here we have a lot of questions from the
46:41
folks in the yeah do you want to keep talking about the future or do you want to answer the questions
46:47
uh let’s go to questions then um the other thing that i want to make sure i notice and i mentioned is
46:55
um the the thing that we’re doing now and that is uh so we
47:02
are getting a lot of interest in pioneer c3 from all kinds of different companies so over the last couple of weeks i just
47:08
had many discussions with people and who basically need help with that and
47:15
also with the other pimc community developers we always thought like oh how
47:20
amazing would it be to work together so now finally it seems like the stars have aligned and
47:25
i’m really excited to announce pimc labs which is a new consultancy that we are
47:31
launching and where we are helping companies to implement or improve their prime c
47:38
models and um basically add this new uh way of doing statistics um
47:45
to to solve analytic problems that are really difficult to solve any other way so if you are uh
47:52
if you want to talk to us send me an email it’s thomas.wiki gmail.com i’d be more than excited to talk to you
47:58
about that and work together with you so that’s my plug you keep telling us all these little
48:03
things i can’t figure out what
48:14
i’m excited for that everyone involved that’s exciting thank you yeah i’m really excited about
48:20
it um i can’t wait for their training content to be honest with you i think that’s going to be some good stuff um
48:26
well how about we get to some questions from uh the audience um l lindsay
48:32
uh was our first question um if you’re familiar with pomegranate uh pop it’s probabilistic modeling in
48:38
python uh could you provide some insight in how pi mc’s three is pymc is different
48:44
uh different use cases different capabilities uh yeah absolutely so pumbra app i think
48:50
is um a package for uh mostly building discrete
48:57
models so um it allows you to build like um i hope i’m not completely mixing this up
49:04
yeah it’s for mixer models uh and hidden markov models so we have discrete states um congruent allows you to
49:12
build these very flexible head markov models and prime c3 is has a much bigger focus
49:19
on continuous models so if you want to build markov models i would say this is a great way
49:25
pi mc3 is focused on continuous models uh however we actually this is another
49:31
future direction is uh brandon willard who is who has done work that i mentioned in all kinds of
49:37
different ways he already um he is working on contributing a new module for pimc3 that focuses on hidden markov
49:44
models and is really flexible and allows you to do that as well so currently with macromoles you need pong grenade or
49:50
something different or something similar but um hopefully soon you’ll be able to do that in time c3 as well
49:59
that’s that’s great um let’s see there was another question from emil dimas uh can you give an
50:06
interpretation of the standard deviation of the standard deviation of the posterior thank you
50:13
yeah so this is always a confusing concept where well we have a parameter right
50:20
that tracks the the standard deviation let’s say we just have normally
50:26
distributed data and we want to estimate the mean and the standard deviation well you can do that in numpy right you
50:32
just compute mean and standard deviation uh and that will give you two numbers so that’s fine
50:37
um in a bayesian framework we are always working with posterior
50:42
distributions and these posterior distributions describe my uncertainty in the graphs themselves so
50:49
i have a posterior over the over the standard deviation of my data and
50:55
of that positive distribution i can also compute the standard deviation so what that would reflect is the
51:01
standard deviation of the standard deviation is my uncertainty in the standard deviation of that parameter
51:07
if i have very few data points my uncertainty in the standard deviation of that on the standard deviation of my
51:13
data will be very wide give a lot of data that will be very narrow so that is what the standard
51:19
deviation of that represents well that was a really good explanation of a hard question i
51:26
think so we have another question from ll lindsay um do you have any
51:32
recommendations for uh tutorials uh references outside the built-in
51:38
tutorials or bayesian methods for hackers um to show real world applications for example
51:44
the rt live um uh example uh that’s a good
51:50
question yeah um so in terms of real world applications hmm well i would say that definitely
51:58
um the the books on the website are pretty good um like
52:05
oswaldo’s book um for actual industry applications
52:10
i don’t have a good answer so that’s definitely one thing that i’m trying to do with my course is like really have those tough um
52:18
application models uh there are definitely some in here that have a real world use
52:25
case like the stochastic volatility model that i showed um and they got some process one for
52:31
example the monolua uh one for co2 levels uh that is like a really good one um
52:38
and um yeah so i i would look at this page but actually it’s a
52:44
good um i don’t have a great answer to that
52:50
fernando erazebal asks do you know when you’ll be releasing your course and approximately how much
52:56
it’ll cost um so i hope
53:02
that we’ll be able to get it together by the end of the year and uh on cost i don’t
53:09
know yet so um that is currently what we’re determining is really the length of the course the thing that i’m
53:17
now thinking is that it will be a shorter course um the first one will
53:23
be very intro level focus just really building an intuitive understanding of the core
53:29
concepts so it’ll be shorter and then there will be another course uh
53:34
for more applied and more advanced topics so rather than one big thing i’m now
53:40
breaking it down into multiple smaller ones and that way i can release it sooner uh at a lower price point per course
53:47
and um and yeah i think it’ll be more targeted
53:52
but thanks for interest great man you seem like a good teacher i want to take that course now
53:58
yeah um so uh we got a question from uh david brochart um which is how has research advanced in
54:05
quantifying uncertainties in uh deep learning inference are there applications embedding pi mc3 inside say
54:12
tensorflow or the other way around building a neural network inside of pi mc3
54:19
yeah that’s a that’s a great question um definitely that’s a very active field of base and deep learning
54:24
and there’s a lot of really interesting things happening there
54:30
it’s not really a focus of pioneer c3 i have some blog posts that show how to do that but there’s
54:37
other packages that are more focused on doing that um and i would say for example tensorflow
54:44
probability has some really strong examples in that domain and so does pyro um i would start looking there
54:52
um and um and actually pi mc4 as well will
54:58
support that so um because prime c4 is built on tensorflow and tensorflow probability uh it can use
55:05
framework like keras and whatnot so the support there will be much much stronger
55:17
amit asks can you give an example or examples of
55:22
how you’ve used this additional information about uncertainty to influence a
55:27
business decision what are your thoughts around conveying uncertainty to management
55:33
that is a really important question and i fortunately just
55:40
uh i don’t know like two years ago maybe found it like the perfect answer and
55:45
it’s based on decision making so the the question is right on right so um
55:52
and this is the story uh that i start with here as well you are a data scientist and your manager
55:58
tells you to like build a model and you come back and you’re all excited about the posterior distribution that you have and he’s like what’s the posterior
56:03
solution right so uh it’s very difficult to communicate these ideas but instead what i’m arguing for
56:11
in this blog post together with robin kumar who’s also doing uh really cool stuff
56:16
and you should follow on twitter is uh that you can use bayesian decision
56:22
making to instead of just providing a posterior distribution that will be the output of
56:27
your model is you use that procedure distribution in the optimization to solve the business problem directly
56:34
so you’re not just saying oh like this is the output now i’m done you’re saying well actually what’s the problem that i’m trying to solve with
56:40
this and here’s the example of um supply chain optimization where we’re
56:47
trying to say how many items do we want to um order from which
56:52
supplier and the idea and the joke is that this is around the operating spaceport so it’s a kind of futuristic
56:58
vibe um and and with that loss function basically the output is not a just like the
57:06
posterior distribution over which supply is how reliable right but actually we’re taking this all the way and solving the problem to tell
57:13
us um directly how many we should order from every supplier so at that point
57:20
right the output is not a posteriority because the output is a decision that we’re making and that decision is
57:27
not just based on the most likely case right but it’s on all possible cases so
57:32
because we’re using the posterior distributions um that uncertainty will be taken into account when we come up with
57:39
that decision so it’s not the decision uh yeah that that solves the most likely case but is the decision
57:46
that’s going to be robust across all possible cases
57:51
well super cool how about we uh thank you for going over just a little bit i don’t know if you’ve lost track of time
57:56
but we are over so we’re gonna do one more question and then we’ll uh we’ll try and wrap it up
58:02
um this question is from abdul rahman uh uh what are
58:09
the key guidelines you will use to decide which toolset bayesian versus statistical models
58:14
uh when you solve a real world problem um well i would um i would say that
58:22
i i view those as similar basin and statistics problems in general so
58:27
i personally never use frequent statistics uh just because i don’t see a reason for that
58:33
but um yeah um i mean i would um i would start with the one that most uh
58:41
where you find the syntax most approachable so yeah i would i would look uh take a look
58:46
at the stand syntax take a look at the prime seats syntax take a look at tensorflow probability
58:51
and and from there basically see what suits your best if you’re a python user i hope that um
58:57
time c3 agrees with you but if not there’s pi stand as well so these are all great packages and um
59:03
and this is really hard to pick among so many great options well that’s great thank you so much uh
59:10
now it’s time for our as we wind up where it’s time for our world famous rant and rave section
59:16
uh where we just have 15 seconds to rant about whatever you want um and uh thomas it’s your soapbox
59:24
all right um i’ve been trying to use tensorflow and find it
59:30
really difficult to use uh whenever i want to do something it’s like there’s five different ways of
59:35
doing that and i never know what the most recent one is and um and if i use the wrong one it’s slow
59:42
it breaks so i just found that to be a challenge and we are facing that challenge too
59:47
with um uh with pime c4 and that’s why i’m so excited
59:53
about jax which is for me much more lightweight and better alternative so that is what i think uh the future is
1:00:02
awesome david what are you ranting about today today today i’ve got a rave usually i’m
1:00:07
good for a rant but i just wanted to reinforce uh and say shout out to the scrub team
1:00:12
on open teams and that includes abraham maxfield daniel alves igor dirk ivano gesara ogasawara jacob
1:00:20
halcyon and tony fast and trent so thank you guys i really appreciate the work well thank you for including me
1:00:27
in the list but uh i guess my rave is uh i love this open source directions
1:00:34
so much because uh we use the word model and real world problems so many times
1:00:41
together if this was deep learning machine learning the real world problem thing doesn’t matter how do i model right but i love i love
1:00:48
hearing the real world applications it makes me wanna uh explore this project a little more um
1:00:53
so that’s all we have time for today uh thank you so much for watching and listening
1:00:59
uh you can follow us on twitter uh at quansiteai um and thomas where can people find you
1:01:04
at nearly ever yeah so i’m on i’m on twitter um kiwiki i’m on
1:01:12
linkedin connect with me there and but yeah twitter is probably the best place to to get in touch uh but yeah thanks
1:01:19
everyone for tuning in thanks for having me on the show it was a blast here well awesome so our open source
1:01:26
army out there uh sign up for open teams today to build your online profile and
1:01:31
uh start keeping track of your contributions and get one of those good remote jobs going well uh take care everybody have a great
1:01:38
day have a great weekend um and thank you all thank you bye