Artwork for podcast Learning Bayesian Statistics
#99 Exploring Quantum Physics with Bayesian Stats, with Chris Ferrie
Teaching Episode 999th February 2024 • Learning Bayesian Statistics • Alexandre Andorra
00:00:00 01:07:31

Share Episode

Shownotes

Proudly sponsored by PyMC Labs, the Bayesian Consultancy. Book a call, or get in touch!

You know I’m a big fan of everything physics. So when I heard that Bayesian stats was especially useful in quantum physics, I had to make an episode about it!

You’ll hear from Chris Ferrie, an Associate Professor at the Centre for Quantum Software and Information of the University of Technology Sydney. Chris also has a foot in industry, as a co-founder of Eigensystems, an Australian start-up with a mission to democratize access to quantum computing. 

Of course, we talked about why Bayesian stats are helpful in quantum physics research, and about the burning challenges in this line of research.

But Chris is also a renowned author — in addition to writing Bayesian Probability for Babies, he is the author of Quantum Physics for Babies and Quantum Bullsh*t: How to Ruin Your Life With Advice from Quantum Physics. So we ended up talking about science communication, science education, and a shocking revelation about Ant Man…

A big thank you to one of my best Patrons, Stefan Lorenz, for recommending me an episode with Chris!

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ !

Thank you to my Patrons for making this episode possible!

Yusuke Saito, Avi Bryant, Ero Carrera, Giuliano Cruz, Tim Gasser, James Wade, Tradd Salvo, William Benton, James Ahloy, Robin Taylor,, Chad Scherrer, Zwelithini Tunyiswa, Bertrand Wilden, James Thompson, Stephen Oates, Gian Luca Di Tanna, Jack Wells, Matthew Maldonado, Ian Costley, Ally Salim, Larry Gill, Ian Moran, Paul Oreto, Colin Caprani, Colin Carroll, Nathaniel Burbank, Michael Osthege, Rémi Louf, Clive Edelsten, Henri Wallen, Hugo Botha, Vinh Nguyen, Marcin Elantkowski, Adam C. Smith, Will Kurt, Andrew Moskowitz, Hector Munoz, Marco Gorelli, Simon Kessell, Bradley Rode, Patrick Kelley, Rick Anderson, Casper de Bruin, Philippe Labonde, Michael Hankin, Cameron Smith, Tomáš Frýda, Ryan Wesslen, Andreas Netti, Riley King, Yoshiyuki Hamajima, Sven De Maeyer, Michael DeCrescenzo, Fergal M, Mason Yahr, Naoya Kanai, Steven Rowland, Aubrey Clayton, Jeannine Sue, Omri Har Shemesh, Scott Anthony Robson, Robert Yolken, Or Duek, Pavel Dusek, Paul Cox, Andreas Kröpelin, Raphaël R, Nicolas Rode, Gabriel Stechschulte, Arkady, Kurt TeKolste, Gergely Juhasz, Marcus Nölke, Maggi Mackintosh, Grant Pezzolesi, Avram Aelony, Joshua Meehl, Javier Sabio, Kristian Higgins, Alex Jones, Gregorio Aguilar, Matt Rosinski, Bart Trudeau, Luis Fonseca, Dante Gates, Matt Niccolls, Maksim Kuznecov, Michael Thomas, Luke Gorrie and Cory Kiser.

Visit https://www.patreon.com/learnbayesstats to unlock exclusive Bayesian swag ;)

Takeaways:

  • Quantum computing has the potential to revolutionize various industries, but it requires specialized tools and education to fully harness its power.
  • Bayesian inference plays a crucial role in understanding and solving problems in quantum physics, particularly in parameter estimation and model building.
  • The field of quantum physics faces challenges in experimental design, data collection, and maintaining the state of isolated quantum systems.
  • There is a need for specialized software that can accommodate the unique constraints and models in quantum physics, allowing for more efficient and accurate analysis.
  • Common misconceptions in quantum physics include the idea of superposition as being in two places at once and the misinterpretation of quantum experiments. Misconceptions in quantum physics and Bayesian probability are common and can be addressed through clear explanations and analogies.
  • Communicating scientific concepts to the general public requires bridging the gap between scientific papers and mainstream media.
  • Simplifying complex topics for young minds involves providing relatable examples, analogies, and categories.
  • Studying mathematics is essential for a deeper understanding of quantum physics and statistics.
  • Taking risks and making mistakes is encouraged in the early stages of a scientific career.

Links from the show:

Transcript

This is an automatic transcript and may therefore contain errors. Please get in touch if you're willing to correct them.

Transcripts

Speaker:

Let me show you how to be a good lazy and

change your predictions You know I'm a big

2

:

fan of everything physics, so when I heard

that Bayesian stats was especially useful

3

:

in quantum physics, I had to make an

episode about it.

4

:

You'll hear from Chris Ferry, an associate

professor at the Center for Quantum

5

:

Software and Information of the University

of Technology, Sydney.

6

:

Chris also has a foot in industry, as a

co-founder of Eigen Systems, an Australian

7

:

startup

8

:

with a mission to democratize access to

quantum computing.

9

:

Of course, we talked about why Bayesian

stats are helpful in quantum physics

10

:

research, and about the burning challenges

in this line of research, but Chris is

11

:

also a renowned author.

12

:

In addition to writing Bayesian

Probability for Babies, he's the author of

13

:

Quantum Physics for Babies and Quantum

Bullshit, How to Ruin Your Life with

14

:

Advice from Quantum Physics.

15

:

So we ended up talking about science

communication, science education, and a

16

:

shocking revelation.

17

:

about Ant-Man.

18

:

A big thank you to one of my best patrons,

Stefan Lawrence, for recommending me an

19

:

episode with Chris.

20

:

This is Learning Asians Statistics,

,:

21

:

Hello my dear Asians, I want to share an

exciting webinar I have coming up on March

22

:

1st with Nathaniel Ford.

23

:

fellow Pimc Cardiff and causal inference

expert.

24

:

In this modeling webinar, Nathaniel will

explore the world of causal inference and

25

:

how propensity scores can be a powerful

tool.

26

:

We will learn how to estimate propensity

scores and use them to tackle selection

27

:

bias in our analysis.

28

:

If that sounds like fun, go to topmate.io

slash Alex underscore and Dora to secure

29

:

your seat.

30

:

And of course, if you're a patron of the

show, you get bonuses.

31

:

submitting questions in advance, early

access to the recordings, etc.

32

:

You are my favorite listeners after all.

33

:

Okay, now back to the show.

34

:

It's Ferry.

35

:

Welcome to Learning Bayesian Statistics.

36

:

Thanks for having me.

37

:

Yeah, thanks a lot for taking the time.

38

:

I'm personally super psyched to have you

on.

39

:

And also, I know a lot of my patrons will

be

40

:

very happy to see you and hear you on the

show because they have asked me for a

41

:

little while now if that was possible to

have you on the show and well apparently

42

:

nothing is impossible in the baysan world

so really thanks a lot for taking the time

43

:

Chris and actually let's start by talking

about what you're doing these days right

44

:

how would you define the work you're doing

nowadays?

45

:

and what are the topics that you're

particularly interested in.

46

:

Sure.

47

:

Yeah.

48

:

So I'm an associate professor at the

University of Technology, Sydney.

49

:

I'm also a co-founder of a tech startup

company.

50

:

And both of these kind of have transformed

me, like at least hopefully temporarily

51

:

into more of a manager than a researcher.

52

:

So the business is developing small,

affordable desktop quantum emulators,

53

:

trying to kind of beef up, enhance, enable

new forms of teaching in quantum

54

:

programming, which doesn't really exist.

55

:

And as a professor, I supervise a handful

of graduate students postdocs.

56

:

I made the mistake, maybe this is like

advice for early career researchers, of

57

:

allowing them all to select their own

projects.

58

:

So I'm supervising students who are all

doing separate projects, all chosen by

59

:

themselves.

60

:

That means that they get to dive deep into

their projects, but I kind of remain at

61

:

the surface level.

62

:

If I'd done it over again, I'd do it

differently with maybe.

63

:

fewer students and working on topics that

really interest me.

64

:

But unfortunately, that doesn't usually

generate much funding because I'm

65

:

interested in the foundations of quantum

physics, and that's more metaphysics or

66

:

you might even say philosophy.

67

:

But it's not bad.

68

:

I get to help young students advance their

careers and learn about new interesting

69

:

topics and there's always time in the

future to eventually settle down.

70

:

Yeah, for sure.

71

:

I didn't know you were also working on an

EdTech company.

72

:

Yeah, you want to tell us a bit more about

that?

73

:

That sounds like fun.

74

:

Well, I'm an elder millennial.

75

:

I was born in the really early eighties,

so that means I have to have side gigs.

76

:

And yeah, it was something that we were

interested in doing at the university

77

:

quantum computing at the university.

78

:

And what I realized was it's a very

abstract thing.

79

:

And it's usually taught from the context

of physics and physics students are happy

80

:

to just be, you know, do what they're

told.

81

:

But computer science students are a little

bit more challenging because they want to

82

:

see something tangible and they want to

build things and see the results of what

83

:

they build.

84

:

So we thought about building this kind of

thing that they can interact with.

85

:

And we made some prototypes and it worked

really well in the context of teaching the

86

:

teaching that I do.

87

:

And we thought, well, and everyone we

talked to in our field about this said

88

:

that they wanted one too.

89

:

And then that kind of led us to the idea

of starting a company.

90

:

So we're at the stage of, of we have, we

have customers, we've built prototypes, we

91

:

have customers, uh, all around the world.

92

:

And, uh, we'll make a big announcement

actually at an event called quantum

93

:

Australia and.

94

:

that will, and then people can pre-order

them, hopefully for shipping later this

95

:

year.

96

:

And it's, so the product is a small

desktop quantum emulator.

97

:

Think about like the relationship between

3D printers that are in classrooms and

98

:

commercial industrial scale 3D printers.

99

:

So our small classroom thing is emulating

the real thing.

100

:

So,

101

:

but it does everything that you need to do

in the context of teaching.

102

:

And it'll come with a full kit to teach

quantum programming to hopefully

103

:

eventually down to the high school and

elementary school levels.

104

:

Nice.

105

:

Yeah, that's super cool.

106

:

And I am going to be honest that I don't

think I can say I know anything about

107

:

quantum computing.

108

:

So why...

109

:

Why would you like to do that?

110

:

What are you, what do you think will that

allow for a better education, basically,

111

:

why would quantum computing help here?

112

:

Well, when we make projections into the

future, we see that we're going to need,

113

:

the quantum industry will need lots of

people, way more people than are in the

114

:

pipeline now.

115

:

So this addresses that market need really.

116

:

So the reason that we want to do it is to

address that market need and do something

117

:

that we think is best fit for it.

118

:

Now as an individual,

119

:

Why would you buy a desktop quantum

emulator and learn about quantum

120

:

programming?

121

:

Well, you know, I think it appeals to the

hobbyists in some sense.

122

:

So if you're someone who buys new tech

stuff on Kickstarter, then you, this is

123

:

the sort of thing that you would buy

because you're curious about it.

124

:

Or maybe you just want to develop new

skills.

125

:

Uh, eventually it will be a subject in, in

high school that students can, can choose

126

:

just like they can choose to do coding now

in high school and programming.

127

:

So quantum computing is something that is,

it's a nascent field, but the 21st century

128

:

will come to be known eventually as the

quantum age, as quantum technologies

129

:

develop.

130

:

Okay.

131

:

And what will that allow us to do?

132

:

I think the only thing I know about

quantum computing is that it's supposed to

133

:

allow you to compute way faster.

134

:

So first of all, the idea I understand

that well,

135

:

And yeah, just can you give us maybe a

rundown on quantum computing?

136

:

Yeah.

137

:

Well, it's not about speed.

138

:

So there are some things that a quantum

computer will be able to do that

139

:

conventional, we call them classical

computers, can't do.

140

:

So the individual steps that occur within

a quantum computer, carrying out an

141

:

instruction is actually slower.

142

:

It's the number

143

:

are way fewer.

144

:

So the device itself is slow, which means

that you wouldn't want to use it for

145

:

simple things like adding numbers.

146

:

Like there's not going to be a quantum

calculator that calculates, that does

147

:

addition faster.

148

:

It's more obscure mathematical problems

that people have connected to real world

149

:

things like applications in cryptography,

in the simulation of chemistry, those

150

:

sorts of things.

151

:

all boil down to these mathematical

problems that are difficult to solve when

152

:

you encode information digitally with ones

and zeros, as you would necessarily have

153

:

to do with your computer.

154

:

If you encode those problems into numbers

that have complex numbers and real numbers

155

:

and negative numbers rather than ones and

zeros, then you can carry out far fewer

156

:

steps to solve your problem.

157

:

And a quantum computer would naturally

encode those numbers.

158

:

and be able to carry out those steps.

159

:

So it's select problems that you would use

this device for.

160

:

It's not just, you know, it's not in the,

it's not this in the faster in the sense

161

:

that eventually we'll have like a iPhone

quantum or something like that.

162

:

It'll be a special purpose component of a

larger computer.

163

:

Just like your CPU outsources graphics

calculations to the GPU, it will outsource

164

:

some quantum

165

:

physics calculations to the QPU in the

future.

166

:

Okay, yeah, yeah.

167

:

Yeah, I see.

168

:

Thanks.

169

:

Much clearer now.

170

:

So, yeah, and I get at least the main

point.

171

:

So, of course, I've already started on

tensions, but I have so many questions for

172

:

you.

173

:

One of my actually planned questions was

that...

174

:

You have a very original origin story

because you claim and you wrote actually

175

:

that quantum physics actually turned you

into a Bajan.

176

:

So tell us why and I'm also curious if

there are any key moments that shifted

177

:

your perspective.

178

:

Right.

179

:

Yeah.

180

:

So yeah, we've been talking about quantum

physics and not Bayesian statistics.

181

:

So it all started when I was a graduate

student and I was interested in this field

182

:

called quantum foundation.

183

:

So it's kind of really trying to

understand the deep underlying questions

184

:

about quantum physics.

185

:

The problem is if you dig deep enough, you

find that quantum physics is just a

186

:

framework built on top of probability

theory.

187

:

You've probably heard of things like the

uncertainty principle, things like that,

188

:

or that quantum physics is a probabilistic

theory.

189

:

And if you look at all of the debates that

happen at the fundamental level and the

190

:

foundational level of the field, they have

more to do with the interpretation of

191

:

probability than they have to do with

physics.

192

:

So when I was a graduate student, I

thought, well, I mean, I'm not going to be

193

:

able to answer these questions until I

understand probability.

194

:

And I suppose in this...

195

:

podcast, I'm preaching to the choir, but I

came out on the other side of that as a

196

:

Bayesian.

197

:

Bayesian, I would put in sort of scare

quotes because I think nowadays you can

198

:

follow the recipes in a book that uses

priors and Bayes' rule and it has the

199

:

title Bayes on it without the need to

actually have an interpretation of

200

:

probability at all.

201

:

So it was more like in order to answer

these questions and have a satisfactory

202

:

understanding of what's going on in

quantum physics,

203

:

You need to have an interpretation of

probability.

204

:

Um, for most physicists, it's just an

implied interpretation that they don't

205

:

really think about.

206

:

But for me, it, you know, it's, it came

out with a subjective interpretation and

207

:

that really helped me understand it.

208

:

Uh, but then I think at some point I was

talking to my thesis committee and they

209

:

didn't like this at all.

210

:

And so most physicists, especially quantum

ones, think probabilities are objective.

211

:

So they told me to do something practical.

212

:

So I transitioned and then tried to start

to apply Bayesian statistics to, you know,

213

:

problems in quantum and quantum physics,

which yeah, they're, it's essentially just

214

:

classical statistics with unfamiliar

models and different loss functions and

215

:

you know, complex numbers are involved in

some sense.

216

:

Um, but yeah, it's basically just a way

to, to derive a likelihood function.

217

:

Now, once you have a likelihood function,

then you're just doing classical

218

:

statistics, it's just a weird likelihood

function.

219

:

Um, so I was able to apply Bayesian

statistics to problems in quantum physics.

220

:

Um, so it was like, I started from this

sort of philosophical point of view and

221

:

then was told to do something practical.

222

:

And so then I was able to.

223

:

some practical things in applying Bayesian

statistics to quantum physics problems.

224

:

Did that change the view that your

supervisors had?

225

:

I think to some extent it did.

226

:

Those techniques and tools that we

developed

227

:

that they're being used in the field,

although it's still dominated with

228

:

frequentist methods.

229

:

Yeah, interesting.

230

:

Yeah.

231

:

In my experience, that's the same.

232

:

So usually people I talk to came to Bass

through practical concern.

233

:

You know, like for instance, a PhD student

who was completely blocked on her paper

234

:

with the classic framework and then she

just tried Bass because while it was...

235

:

of her last resort and it solved all of

her problems and now she's just doing

236

:

that.

237

:

But that's a very practical motivation.

238

:

And yeah, I see most people coming from

that angle.

239

:

You're actually more in the outlier side

where you've been more interested in the

240

:

epistemological point of view and then

shifted to actually doing it.

241

:

And yeah, actually what I've

242

:

It's actually useful.

243

:

Just show them.

244

:

And then they'll be like, yeah, that does

look good.

245

:

And that does solve the problem we were

having.

246

:

So why not try that?

247

:

So in my experience, that's been the same,

too.

248

:

And I'm curious, when was that work you

did on practical Bayesian inference?

249

:

When did you do that?

250

:

Oh, that's gotta be 16.

251

:

Yeah.

252

:

12, 16 years ago.

253

:

And we, so it kind of culminated in, we

built this tool, we call it Qinfer, and

254

:

it's basically a sequential Monte Carlo

integrator that just naturally was able to

255

:

solve the kinds of problems that people

have in quantum

256

:

Because it's quite difficult actually to

use standard tools.

257

:

Often they don't play nice with complex

numbers and things like that.

258

:

Don't naturally have the kind of loss

functions and things that we use in

259

:

quantum physics, kind of matrix

manipulations that we have to do.

260

:

And at the time there wasn't that many,

right?

261

:

Computation-based statistics is a

relatively new thing.

262

:

There was a few tools, but not many.

263

:

And so we ended up building our own and

it's been used many times over the years.

264

:

And that was maybe 10 years ago.

265

:

I stepped back from that and handed it off

to the next graduate student.

266

:

Yeah, that's why I asked you, when did you

do that?

267

:

Because just a few years ago, there wasn't

a lot of tools to do that.

268

:

So yeah, like you had, I'm guessing you

had to write the algorithm from, from top

269

:

to finish on your own.

270

:

Yeah.

271

:

And honestly, sometimes that's, that's

better to do it that way.

272

:

I mean, if you want to really deeply

understand something, you have to build it

273

:

yourself.

274

:

You know, we can't build everything from

scratch.

275

:

I mean, if, if you want to understand

particle physics, you can't go build your

276

:

own particle collider, but, uh, for things

that you, you have the capacity to build,

277

:

I would always recommend building it

yourself or at least attempt to, and then

278

:

realize what all of the problems, uh, are

going to be if you wanted to make a really

279

:

slick product.

280

:

So get it to the point where you've built

a prototype and then you really kind of

281

:

deep start to deeply understand.

282

:

what's going on because a lot of times,

especially with really usable products,

283

:

they're really slick and they're just

black boxes.

284

:

And yeah, you can push the buttons and use

them, but you don't end up developing a

285

:

deep understanding of, of what's going on.

286

:

Yeah, yeah, for sure.

287

:

Even though hopefully if you had to do

that today, that would be easier.

288

:

You could use building blocks instead of

really just starting from scratch.

289

:

And thankfully- Well, I mean, an example

is I...

290

:

Yeah, I can give you an example.

291

:

So I have a student, an undergraduate

student that I suggested trying a new

292

:

it's jargon, but I'm sure people have

heard about it.

293

:

Maybe you heard about it.

294

:

The Stein variational gradient descent

method, which is a deterministic

295

:

integration method and, you know, it's

built into, um, Pi MC.

296

:

Uh, so I, the student can go and can go

and try that, although it is quite, it's

297

:

still quite difficult for them to build,

build the quantum mechanical models that

298

:

they have to build.

299

:

So first I have them do it from scratch.

300

:

And, uh, of course it

301

:

It works to some extent, but it's not very

efficient.

302

:

There are a lot of things that tricks that

come up in numerics.

303

:

Like, what do you do if you're trying to

take a logarithm and there's something

304

:

close to zero, right?

305

:

Then you don't want them to have to figure

out all those things.

306

:

Have them build it first and then go.

307

:

Yeah, yeah.

308

:

Yeah, basically using...

309

:

Yeah, I like that.

310

:

Basically using a version from scratch

that's...

311

:

Simplified and then when you need to go

industrialize that, well, just use the

312

:

tools you have already on the shelf and

maybe customize them if need.

313

:

That's the beauty of.mc where you building

blocks basically that you can personalize

314

:

into your own Lego construction in a way.

315

:

Yeah, for sure.

316

:

But that's awesome.

317

:

Well done on doing that thing.

318

:

And were you already using Python at the

time, 16 years ago, when you were doing

319

:

your own SMC or was it something else?

320

:

No, the first version was built in Matlab,

but as you might anticipate, we ran into

321

:

license issues when we ended up using

every one of the entire university's

322

:

global optimization toolbox licenses.

323

:

And so then we thought, well, this is

silly.

324

:

So then we moved over to Python.

325

:

The first one, yeah, it was kind of like

the transition.

326

:

So we had an early version built in 2.7,

and then we moved to 3.

327

:

Nice.

328

:

Yeah.

329

:

That's really fun.

330

:

Yeah, in SMC, I know there are also some,

like you can do that here with PMC now.

331

:

So yeah, if one of your students is

interested,

332

:

They can contact me and I'll direct them

to the persons who like doing that on the,

333

:

on the PIMC community.

334

:

And, and you personally, do you have any

specific instances to share or insights

335

:

that you gained by adopting a Bayesian

approach in your, in your research?

336

:

I mean, it's hard to know, I suppose.

337

:

I mean, I haven't given it a lot of

thought, right?

338

:

Because it wasn't like I had this problem

and classical techniques weren't working

339

:

for me.

340

:

And then I switched over and found, you

know, a particular set of Bayesian

341

:

techniques that ended up working.

342

:

I recommend it to people because a lot of

times, especially when you're thinking

343

:

about things deeply and foundationally,

like...

344

:

You know, what are these things mean in

quantum physics?

345

:

Um, it, I always go back to simple

classical examples and say, if you can

346

:

understand this, or I guess it's a more

negative thing, like if you can't

347

:

understand this, then you're not going to

even have a chance at understanding the

348

:

more complicated thing.

349

:

So, you know, I go back to coin tosses and

I say, okay, what does it mean in the

350

:

context of a coin toss?

351

:

And if you don't understand it there,

you're not going to understand that

352

:

quantum version of it.

353

:

And the, yeah, the subjective

interpretation of

354

:

of probability just makes things more

natural.

355

:

I mean, it gives you a framework for

thinking about things that you can always

356

:

build on rather than the classical

approach, which it doesn't give you that

357

:

framework at all.

358

:

It's just grasping at straws and saying,

okay, you know, what recipes work in this

359

:

situation?

360

:

And there isn't one coherent framework

sitting behind it.

361

:

Whereas the subjective interpretation

gives you that.

362

:

And so you might not, yeah, you might not

363

:

It's not like it gives you a specific set

of tools that you can apply in every

364

:

situation, but it gives you that footing,

that foundation that you can build upon

365

:

and always have that level of comfort,

philosophical comfort saying, I

366

:

understand, I know what's going on.

367

:

Yeah, for sure.

368

:

And to build on that question, do you have

a favorite study or paper of yours where

369

:

you used some Bayesian stuff at one point?

370

:

I'm curious to see, and I'm guessing

listeners too, curious to see where

371

:

Bayesian stats is useful when you do

research in quantum physics.

372

:

Yeah, there's lots of papers.

373

:

I think most of them would be readable for

someone coming from Bayesian statistics

374

:

without knowledge of quantum physics.

375

:

Because again, I try to frame it in this

way where the quantum physics, the only

376

:

point of the quantum physics is to arrive

at the likelihood function.

377

:

And once you have that, then you can just

do all the things that you're used to

378

:

doing.

379

:

Is it because your likelihood functions

are always extremely exotic?

380

:

Yeah, so the standard simple quantum

experiment would be about estimating the

381

:

parameter in a multinomial distribution.

382

:

So you can think of a quantum experiment

as rolling a die and trying to estimate

383

:

the probabilities for the faces of the

die.

384

:

Yeah, but...

385

:

The thing is we don't, um, the, we have

like loss functions that, I mean, yeah,

386

:

they, there's some, some major season

things in there.

387

:

And then the issue is like, we have these

loss functions that aren't, aren't ever

388

:

used in, in classical statistics.

389

:

And so a lot of the results, uh, just

don't apply.

390

:

So you, you know, you, you can, you can

sometimes appeal to, uh, like the law of

391

:

large numbers or, or some of these

392

:

you know, these theorems, but they,

strictly speaking, our models don't really

393

:

adhere to those, the assumptions that go

into those theorems.

394

:

So not only do we have weird loss

functions, that allowed probabilities for

395

:

the faces of the die are constrained in a

weird way that relates to a positivity of

396

:

some matrix that sits down the pipeline.

397

:

So it's, yeah.

398

:

So oftentimes you would, you would, if you

did it naively, you would end up

399

:

estimating, um, things that make

probabilities negative, which obviously

400

:

doesn't make sense.

401

:

So, um, yeah, there's weird constraints.

402

:

There's an atypical, um, statistical

models and, and then the loss functions

403

:

that we use are quite different.

404

:

So, but, you know, if, if you know enough

statistics and, and

405

:

can accept that there are different, you

know, that the possibilities extend beyond

406

:

what you're used to, then yeah, you can

you can work with it.

407

:

A lot of times the things that you'd

naturally try don't work.

408

:

But, you know, it is still just a

classical statistical problem.

409

:

We there was there was one paper where we

were trying to find another way to

410

:

problem in parameter estimation in quantum

physics is the parameter that you're

411

:

trying to estimate is itself a matrix.

412

:

So it's not a real value.

413

:

It's not a real value vector.

414

:

It's a complex value matrix.

415

:

And that's the thing you're trying to

estimate.

416

:

So I don't know if you're doing density

estimation, that sort of thing.

417

:

It's similar to that.

418

:

But we wanted to find the Bayes estimator

for a particular loss function that

419

:

involves square roots of

420

:

And if you assume that all the matrices

are diagonal, then you're back to a

421

:

classical statistical problem and you end

up with this funny loss function for

422

:

classical probabilities that's somewhat

related to some loss functions that are

423

:

used in learning theory.

424

:

And then we said, oh, well, people

actually haven't found the Bayes estimator

425

:

or, let's just say, the minimax estimator

for that particular function.

426

:

So our quantum result immediately implied

a result just that was purely classical.

427

:

And we, the papers titled the papers

estimating the bias of a noisy coin.

428

:

So it's, uh, this, this actually crops up

in, in social, uh, some social studies.

429

:

So if I, if I ask you, if you cheat on

your taxes, you're going to say no.

430

:

So how do they do the sampling?

431

:

What they do is they, they introduce some

randomness.

432

:

So they they'll say, okay.

433

:

roll a die, if the die comes up one, say

yes no matter what.

434

:

And so that the person who says yes can

always claim that the die came up one.

435

:

And so they feel like they can be honest.

436

:

But if that probability of people cheating

is really low, then you might get only one

437

:

or two people saying yes, but one in six

times they were supposed to say yes

438

:

anyway.

439

:

So if you just naively kind of

440

:

did methods of moments or some linear

inversion, you would come up with negative

441

:

probabilities.

442

:

So this is exactly a problem that's

embedded in a quantum mechanical problem.

443

:

And so sometimes there's some nice overlap

there.

444

:

Yeah, for sure.

445

:

That sounds like fun.

446

:

And for sure, if you can add these papers

to the show notes, please do, because I'm

447

:

pretty sure listeners are going to be

happy to.

448

:

to check those out.

449

:

I already put some cool links in the show

notes for people, but definitely papers

450

:

are always appreciated, so feel free to do

that.

451

:

This is a safe place where we can all

share our love for academic papers.

452

:

Great.

453

:

Yeah, I should warn the listeners though.

454

:

Yeah, a lot of them are, they're cavalier,

like a typical physicist.

455

:

So it's very...

456

:

We often take a conceptual approach to

these things.

457

:

Okay, interesting.

458

:

Well, I read it because it must be pretty

different from a statistics paper.

459

:

I don't think I've ever read a quantum

physics paper.

460

:

So yeah, for sure.

461

:

I think I'm going to start by your books

though, your books for children.

462

:

I'm embarrassed to say, I think I'm going

to learn a lot from them.

463

:

So I'm going to start by getting to walk

my way up to your papers.

464

:

Sounds much, much clearer.

465

:

And maybe before actually talking a bit

more about quantum physics and what you do

466

:

and also the work you do on your

children's books, but also science

467

:

communication in general, and I'd like to

keep talking a bit more about Bayesian

468

:

stats because I'm curious, I'm always

curious when I talk to a practitioner like

469

:

you and so someone who is not...

470

:

by training a statistician, but someone

who really uses Bayesian statistics for

471

:

their area of expertise.

472

:

What do you see as the biggest pain points

in the Bayesian workflow right now?

473

:

I think, as I mentioned before, the

software that is typically used off the

474

:

shelf doesn't accommodate the quirks and

things that come up in quantum models.

475

:

Some of them, they just won't accept

complex numbers, for example.

476

:

When I first attempted to use TensorFlow

way back, TensorFlow 1, you couldn't even

477

:

use complex numbers.

478

:

to go back to the source code.

479

:

And at that point, you might as well just

build it yourself.

480

:

So yeah, complex numbers, matrix

manipulations, we often have, as I said,

481

:

lots of constraints.

482

:

And when you attempt to use something out

of the box, if it works at all, your whole

483

:

screen is filled with warnings.

484

:

And it isn't.

485

:

It isn't as nice as the demos of the

software.

486

:

So I think for me, and possibly for people

that are running models with lots of

487

:

constraints, this is the biggest pain

point at the moment.

488

:

Obviously, the software will accommodate

constraints, but it doesn't.

489

:

It doesn't seem to do so in a way that's

natural and easy.

490

:

Yeah.

491

:

So ideally that like in an ideal world,

that would be what you'd like to see to

492

:

help adoption of patient training.

493

:

Yeah.

494

:

I mean, like a really concrete example

would be, you know, I want to do

495

:

sequential Monte Carlo on some simple

estimates.

496

:

I'm doing an experiment where I roll a die

several times and I want to estimate the

497

:

probabilities.

498

:

It's of some biased die, but the

probabilities come with a long list of

499

:

linear constraints.

500

:

So not any probability will do.

501

:

When you're doing the resampling, what is

it that the software is doing to

502

:

accommodate those constraints?

503

:

approach is like, what doesn't really

matter because there is no constraints.

504

:

And so you can just throw a Gaussian on it

and you know, it, nothing.

505

:

Yeah, it's simple, but when you have these

constraints, um, yeah, it makes, it makes

506

:

things far, far more challenging.

507

:

And sometimes the software just doesn't,

doesn't accommodate those.

508

:

Yeah, yeah, no, for sure.

509

:

I understand your pain.

510

:

And I'd like to make your wish come true,

but that's a hard one because in here,

511

:

you're hitting a limitation, I would say,

of the development process where you have

512

:

to choose at some point if your package is

going to be general or specific.

513

:

And packages like Stan, Climacy,

TensorFlow, they have to be general

514

:

because they are adopted by so many people

with so many different backgrounds and so

515

:

many different uses that we have to make

choices that are going to work for most

516

:

people and that are going to be optimal

for most use cases.

517

:

But that means for sure it's like

518

:

If you're trying to accommodate everybody,

nobody's going to be accommodated

519

:

perfectly.

520

:

Right.

521

:

So, yeah, like it seems to me like someone

should go there and basically build a

522

:

package on top of PIMC that just like

addresses what you folks pain points are

523

:

in quantum physics.

524

:

Basically.

525

:

I know there is such a package for

astrophysicists.

526

:

Of course, I don't remember the package

name right now, but I'll try to remember

527

:

and put that in the show notes.

528

:

And I know that package built on top of

times is really, really used a lot in the

529

:

astrophysics field.

530

:

I'm not aware of any package like that in

the quantum physics realm.

531

:

But if any listeners do, but then please

reach out to me and I'll pass that on to

532

:

Chris.

533

:

I'm sure his PhD students are going to be

grateful.

534

:

Yeah.

535

:

Or if anybody wants to do that, get in

contact with Chris, I'm sure he would have

536

:

valuable points for you about what he'd

like to see in particular.

537

:

I think it's honestly there's a research

question in there as well, right?

538

:

At least when we were doing it, that

particular method that we were using, it

539

:

was never applied or developed in the

context of constraints.

540

:

And so what you do when you're faced with

constraints, at the time anyway, it was

541

:

like sort of an open research question.

542

:

So yeah, it's fair that...

543

:

It's fair that the software just doesn't

solve it for you because it may not be a

544

:

there may not be an actual solution yet.

545

:

Yeah, that's a good point also.

546

:

And so now I'd like to ask you a bit more

about quantum physics per se, because,

547

:

well, I'm always very curious about

physics.

548

:

So what in your line of research, what are

the biggest questions, the biggest

549

:

challenging you face currently?

550

:

So we're at this weird transition point in

the field of quantum technology where we

551

:

can't in laboratories, university

laboratories, build bigger devices.

552

:

So we kind of count the power of a quantum

computer in the number of quantum bits or

553

:

qubits that we can control.

554

:

And nowadays it's very easy to get one

qubit.

555

:

was very difficult, but now there are many

different modalities, trapping atoms,

556

:

using states of light.

557

:

All of these sorts of things can now be

used to encode a single qubit, and that

558

:

can be done in the standard physics lab.

559

:

Going beyond that becomes more difficult

and you need much more funding to do it,

560

:

but going much further beyond that is not

a possibility within an academic.

561

:

context.

562

:

And so you need some large government

organization or collaboration to do it, or

563

:

you need industry to take over.

564

:

So we're at that cusp where the largest

devices are ones that are being developed

565

:

by companies, companies like IBM, Google,

startup companies like Rigetti, IonQ.

566

:

There's a whole host of them now.

567

:

And what they're doing, obviously,

568

:

secret now.

569

:

So it's a weird place to be.

570

:

I can't tell you, I can make guesses about

where they are, what they're doing, what

571

:

their problems are.

572

:

But if they wanted my help, I'd have to

sign an NDA, or they'd have to pay me and

573

:

I wouldn't be able to tell you.

574

:

So we've kind of transitioned into

575

:

We're moving out of university research

labs into government and company and

576

:

multinational company R&D labs.

577

:

They have the same problems, but at a

larger scale that university researchers

578

:

had, which is just that to maintain the

state of an isolated quantum system is

579

:

very difficult.

580

:

Any interaction.

581

:

cosmic ray that comes in that you

obviously can't control will degrade the

582

:

information that's being encoded in these

systems.

583

:

And so they're very fragile.

584

:

We need to work out ways to provide better

isolation, but complete isolation is not

585

:

good either because you have to control

them to carry out the instructions that

586

:

you want.

587

:

So it's kind of this Catch-22 where you

want it to be completely isolated from

588

:

everything except for when you want to

actually.

589

:

go in there and manipulate it in some way.

590

:

So yeah, these are the problems.

591

:

And I think theoretically there's still

that big question about can it even be

592

:

done?

593

:

Can we even build a quantum computer?

594

:

There doesn't seem to be a reason why.

595

:

If it turns out that we couldn't, we'd

learn a lot about the nature of reality

596

:

and the reason for why that's the case.

597

:

But

598

:

I think have the potential to be answered

in my lifetime.

599

:

Can we build a large scale fault tolerant

error corrected quantum computer that

600

:

carries out some calculation that would

have been impossible to carry out with

601

:

digital electronics?

602

:

Yeah, yeah, that's pretty fascinating.

603

:

And I'm really impressed by the depth and

the width.

604

:

of topics in the research of physics.

605

:

It's just incredible.

606

:

I would refer to listeners to episode 93

that I did at CERN, the summer, I mean,

607

:

2023 summer, where we went deep on what do

they do at CERN, what type of research,

608

:

what does that mean, why even do that.

609

:

And you'll see, well, some, you know,

cross topics with what Chris is talking

610

:

about, but also things that are completely

different.

611

:

And that's just incredible to see how wide

these fields are.

612

:

And that sounds to me that's pretty

incredible because in the end, that's

613

:

just, you know, trying to understand the

universe.

614

:

So it's kind of doing the same thing, but

it brings you...

615

:

to directions that are completely,

completely different.

616

:

And that's really the funny, one of the

fascinating things, I think, of these

617

:

topics.

618

:

And of course, go to the video version of

the episode 93.

619

:

You have the audio version if you have,

but that was a video documentary inside

620

:

CERN.

621

:

So I highly recommend checking out the

YouTube link that I will put in the show

622

:

notes.

623

:

And actually, I'm curious, Chris, about

also because now, as you were saying, you

624

:

kind of have a management role, which

implies thinking a lot about the future.

625

:

So I'm wondering, where do you see the

field of quantum mechanics headed in the

626

:

next decade?

627

:

Also, maybe how do you see patient stats

still helping in this endeavor?

628

:

That's a good question.

629

:

I think much like astronomy, for example,

Bayesian techniques will see a wider

630

:

adoption because at the moment, the way

that a laboratory quantum physics

631

:

experiment happens is really foreign to

someone who does machine learning or data

632

:

science where you have some data set and

then you need to analyze it.

633

:

No, what they do in labs in physics

departments is if the data isn't what you

634

:

wanted, then you just throw it out and

start again.

635

:

And, and you work until you have like

really clean data sets.

636

:

So all of the all of the problems with

data sets and things like that don't

637

:

happen in physics labs.

638

:

The physicists want to see the answer in

their data.

639

:

So the really sort of data scarce regime

is unacceptable to them.

640

:

They need to see it on an oscilloscope or

something.

641

:

The probability distributions essentially

have to be delta functions for them before

642

:

they accept that the experiment actually

worked.

643

:

But that's because we're doing really

small-scale experiments.

644

:

Once those experiments grow and become

large, we won't be able to do that

645

:

anymore.

646

:

If an experiment takes a week to run,

647

:

You're not going to say, do it over again

until you see a nicer data.

648

:

You're just going to have to accept that

that's the data set and you have to, you

649

:

know, get as much information out of it as

possible.

650

:

And that's going to require utilizing the

assumptions that you're making.

651

:

In a sensible way, which will lead you to

sort of Bayesian techniques.

652

:

So I think we will see wider and wider

adoption within the quantum research

653

:

fields.

654

:

of Bayesian techniques going into the

future, much like we have in the last two

655

:

decades in astronomy.

656

:

Hmm.

657

:

Yeah.

658

:

Uh-oh.

659

:

Yeah, fascinating and...

660

:

I really hope that these big questions you

were talking about are going to be

661

:

answered, at least some of them, because

I'm just so curious about that.

662

:

That would be just fascinating to have

some of these answers at least come our

663

:

way in the coming years.

664

:

um, relativity in quantum physics and how

you can merge that.

665

:

And so that's definitely would be

incredible to at least understand that a

666

:

bit better.

667

:

And also, and I'm also fascinated by the

fact that how do you do the experiments on

668

:

this realm for now is just super

complicated.

669

:

Yeah, I think those are huge questions.

670

:

I don't even think we've really formulated

the questions correctly.

671

:

I mean, that's my take on it.

672

:

We have a theory that works really well at

the moment.

673

:

In every regime we can test, our current

best model quantum field theory works

674

:

incredibly well.

675

:

It's places that we don't even understand

like inside the event horizon of a black

676

:

hole.

677

:

in principle, we can't even go there to

get the data that we would need to find

678

:

out if the theory works there.

679

:

There's various takes on it.

680

:

It's just a pessimistic take, which is

like, maybe we've hit the limits of what

681

:

we can understand given our capabilities

in the universe.

682

:

And then, yeah, a more positive view is

like, well, eventually someone will come

683

:

up with some idea

684

:

there was something that nobody could have

seen coming.

685

:

That's typically how paradigm shifts have

worked in the past.

686

:

So there's no reason to think

pessimistically that will stop.

687

:

But who knows, it might be the case.

688

:

Yeah.

689

:

I mean, I do hope for the second option,

but you can never know.

690

:

And actually now

691

:

I love the fact that you do a lot of

science communication, of course it's also

692

:

a job of these podcasts, so it's always

something I'm very happy to talk about and

693

:

I'm wondering if there are some common

misconceptions you've seen about quantum

694

:

physics, maybe even about

695

:

Oh, yeah.

696

:

Well, I wrote an entire book for, not for

children.

697

:

It's, yeah, you may have to edit this part

out because the book's called Quantum

698

:

Bullshit.

699

:

I don't know if that's allowed in the

podcast.

700

:

I'm French, so we have no worries with

swear words.

701

:

Yeah, in Australia it's similar.

702

:

Yeah, so that's the title of the book.

703

:

The subtitle is kind of a science comedy.

704

:

So the subtitle is How to Ruin Your Life

with advice from quantum physics.

705

:

And it kind of goes through a lot of the

common misconceptions and how each of

706

:

these major concepts in quantum physics

are misused.

707

:

Things like superposition, entanglement,

quantum energy, quantum uncertainty, these

708

:

sorts of things, how they typically are

misused.

709

:

And yeah, what's the most sensible kind of

way to understand them without having the

710

:

mathematical background that underpins the

framework of the theory?

711

:

So yeah, there's lots of them.

712

:

And if you want the comprehensive list,

definitely check out the book.

713

:

I'll give you like a typical

714

:

means things can be in two places at once.

715

:

And that just like, just saying it out

loud should make it clear that that's a

716

:

logical contradiction.

717

:

Because, you know, a dichotomy between

true and false, and you can't have

718

:

something that's both true and false.

719

:

So sort of a logical contradiction.

720

:

But that being said, you still, you know,

physicists will still say things

721

:

that sound kind of like that.

722

:

So an example might be this famous double

slit experiment where you have some sort

723

:

of screen, it has two holes in it, and you

fire electrons at it and you see an

724

:

interference pattern on the other side

instead of just two dots where the

725

:

electrons landed, suggesting that the

particles interfere with each other.

726

:

And if you do it one particle at a time,

that means it has to interfere with

727

:

itself.

728

:

which means it had to have gone through

both slits at the same time.

729

:

So the electron had, or whatever particle

it is, had to be in both of those places

730

:

at the same time.

731

:

But we always run into these problems when

we try to explain what's going on in

732

:

quantum physics by analogy to our everyday

world.

733

:

It's just a different world that we don't

have access to.

734

:

We don't have a language and a familiarity

with.

735

:

So we have to use these analogies.

736

:

But...

737

:

you know, they very quickly break down.

738

:

So that's absolutely not what's happening.

739

:

Uh, and things can't be in two places at

once.

740

:

And yeah, you shouldn't, uh, you should

buy a quantum crystal or something because

741

:

it promises that, that it can do that.

742

:

And for the Bayesian, I find actually, um,

uh, yeah.

743

:

So, you know, when you

744

:

You can kind of explain to people the way

I do it now is to walk through that idea

745

:

that in quantum physics we have these

concepts and we have to use a language

746

:

that we're familiar with but that language

isn't really suited for trying to do

747

:

anything beyond explain that one special

thing.

748

:

You can't extrapolate using those

analogies because you'll quickly fall prey

749

:

to misconceptions.

750

:

So

751

:

That's typically how I explain it in the

context of quantum physics.

752

:

And quantum physics is actually quite

popular in the popular culture.

753

:

I don't find that Bayesian probability is

so popular in popular culture.

754

:

So, you know, the word quantum crops up

all the time, attached to things.

755

:

Nobody's selling Bayesian healing

crystals.

756

:

So, these aren't like popular.

757

:

Oh, that's actually not a bad idea.

758

:

Yeah.

759

:

But so you don't need to approach it the

same way because you're not typically

760

:

talking to a lay audience when you're

talking about misconceptions and Bayesian

761

:

probability.

762

:

Usually it's someone technically minded

who knows something about some technical

763

:

topic that the probability is being

applied to or probability itself.

764

:

In physics, the main problem that people

have, you could call it a misconception,

765

:

is that Bayesian methods are subjective,

whereas frequentist methods are objective.

766

:

And as a scientist, you need to strive for

objectivity.

767

:

So that means that you shouldn't use

Bayesian methods and you have to use

768

:

frequentist methods.

769

:

But the easy thing to point out is to...

770

:

What you could do is just...

771

:

have them walk through how they would

apply frequentist methods and then point

772

:

out that they had options and then they

made their subjective judgments on which

773

:

options they were going to choose to solve

their problem.

774

:

So it's no less subjective.

775

:

And in some sense, it's worse in the sense

that you're not being honest about the

776

:

biases that are going into what you're

doing.

777

:

So yes, Bayesian methods are absolutely

subjective, but they're subjective in the

778

:

most honest way possible.

779

:

Yeah, that's usually the way I go about it

also.

780

:

The faster you're going to abandon the

idea that there is an objective way of

781

:

seeing reality, at least the way we are

made, you know, if you're homo sapiens,

782

:

the faster you'll be able to think about

real ways to actually try to understand

783

:

what's going on.

784

:

And so, yeah.

785

:

It's usually the way I go about it.

786

:

But yeah, I mean, these are fascinating

topics.

787

:

I, we've actually covered some of them in

some of the episodes we've already done on

788

:

the show.

789

:

So the one, one before you was episode 97

with Alien Downey where he actually talked

790

:

about that where.

791

:

He has also a blog post about it comparing

this idea that Bayesian results converge

792

:

to the frequentist results to the limit.

793

:

And so that was interesting to talk about

it with him because he actually argues

794

:

that it's never the same.

795

:

And that's not a problem.

796

:

You should still choose the Bayesian

framework, actually.

797

:

But that was interesting.

798

:

So you have that for people interested and

also I'll put in the show notes.

799

:

So I'll put that one and I'll put in the

show notes, episode 50 and 51.

800

:

50 was with Aubrey Clayton, who wrote an

amazing book called Bernoulli's Fantasy

801

:

and the Crisis of Modern Science.

802

:

So that's more about the history of

statistics and how basically, how and why

803

:

came to dominate the scientific world.

804

:

So much more epistemological, very, very

fascinating book.

805

:

And episode 51 with Sir, only Sir we've

had on the podcast, I think, Sir David

806

:

Spiegelhalter about risk communication,

how to talk about risk, especially to a

807

:

lay audience.

808

:

and people who are not educated in stats

or in the scientific method.

809

:

And that was, that was way closer to the

COVID pandemic.

810

:

So that was very interesting to talk about

that with him, because these topics were

811

:

absolutely important in time of pandemic

or very stressful situations.

812

:

Right.

813

:

Who would think so, right?

814

:

That the nerds actually had tried all

along to talk about stats and

815

:

probabilities.

816

:

This can save you during a pandemic.

817

:

But yeah, I mean, this is also something

that I think must be added in these

818

:

discussions.

819

:

Often, it's not really in the papers that

you see these misconceptions, but it's

820

:

more in the way the papers are interpreted

by people who are not equipped to read the

821

:

papers.

822

:

And so I think there is a...

823

:

a job in the world that needs to be

filled, which is basically making the

824

:

bridge between scientific papers and then

what ends up in the newspapers.

825

:

And that is a bridge that still has to be

built.

826

:

And we're trying to do that in a way with

our work, but it's still so much things to

827

:

do still.

828

:

Sometimes my game is really to do that.

829

:

It's trying to see what people are talking

about on Instagram or stuff like that.

830

:

And then actually try and go to the source

that they are supposed to quote, you know,

831

:

to site.

832

:

And then you see that basically it's just

like the first person who reported on the

833

:

paper did understand the paper or just

read the abstract and the title.

834

:

And then just everybody cite that first

source.

835

:

So basically the first error is just like

trickled down and that's just fascinating.

836

:

Yeah.

837

:

Yeah, I think the solution has to sort of

include actually that people write fewer

838

:

papers.

839

:

I mean, there's over a million academic

journal articles published every year, and

840

:

that's more than we can read, right?

841

:

But there's the perverse incentives in

academia now that kind of force you to do

842

:

this, which means also that like most of

those papers shouldn't have been written,

843

:

I think it would be better if we had a

more careful approach where the result is

844

:

fewer papers that are better written.

845

:

Yeah, that could have been more.

846

:

And also it's something we've talked about

on the podcast several times, incentives

847

:

in academia.

848

:

It's hard to change, but needs to be

changed.

849

:

But yeah, hopefully that will...

850

:

And having people like you in academia

definitely helps.

851

:

Well, hopefully with time, it's going to

evolve.

852

:

But yeah, and we could continue on that

road, but it's going to be a three-hours

853

:

episode, and I don't want to take too much

time to you.

854

:

And actually, that's a very, it's the very

first episode that we do where we are

855

:

actually time traveling, right?

856

:

Because it's still January 15 for me.

857

:

at night and it is January 16th in the

morning for Chris.

858

:

So thank you for calling from the future,

Chris.

859

:

We solved the glass problem.

860

:

The sun rises tomorrow.

861

:

Yeah, I can tell you that.

862

:

Yeah, I can see for now, no apocalypse.

863

:

So that's cool.

864

:

Glad about that.

865

:

Yeah, I had other things to add about your

very good points about communication and

866

:

so on.

867

:

But of course I...

868

:

I think I forgot about them.

869

:

I will just refer people to the show

notes.

870

:

I'm gonna put the episodes I mentioned in

there.

871

:

And oh yeah, one thing, I tracked down the

Python package I was talking about for

872

:

Astrophysics.

873

:

So the package is actually called

Exoplanet.

874

:

And yeah, it's a package that's built on

top of PymC.

875

:

to do probabilistic modeling of time

series data in astronomy with a focus on

876

:

observations of exoplanets.

877

:

So I put the notes, the link already in

the show notes, and that's developed

878

:

mainly by Dan, Ferm, and Mackey.

879

:

So people who are working on that

definitely take a look at a very cool

880

:

package, very well maintained.

881

:

So Chris.

882

:

I've already taken a lot of time from you,

but I'm curious.

883

:

I want to talk a bit about your children's

book.

884

:

Of course, you've written about quantum

physics, about general relativity.

885

:

Patient statistics also, you've written a

book, I think, about that.

886

:

First, I'm definitely going to buy those

books if one day I have kids.

887

:

That's for sure.

888

:

I'm not going to read them stories

about...

889

:

crystals and things like that, much more

about that kind of thing.

890

:

No, first, keening aside that I think

that's a very good service you're making

891

:

because definitely there is a big lack of

scientific culture, I would say in

892

:

general, in the audience, just

understanding probability.

893

:

The main thing I have to face is often

things like

894

:

Well, you said that thing would happen

with a 30% chance.

895

:

It didn't happen.

896

:

Hence the model was wrong.

897

:

And that's just like, this is kind of the,

this part of the misconceptions on, on the

898

:

part of, this is the burden of a

statistician.

899

:

But I think it's extremely important to

make people more aware of the scientific

900

:

methods, more scientific savvy.

901

:

First pick is way more interesting than

what pop culture makes it look like.

902

:

You know, you don't have to be crazy.

903

:

You don't have to wear a white coat.

904

:

You don't have to be a genius to

understand science.

905

:

And you don't have to be a genius to use

science.

906

:

So, yeah, I think it's extremely important

what you're doing.

907

:

And mainly to go to my question, how, how

do you approach simply

908

:

simplifying such complex topics for young

minds and yeah, how do you think about the

909

:

way you teach that?

910

:

Yeah, that's a good question.

911

:

I think you hit on a lot of good points.

912

:

And there's a lot of obvious traps that

people fall into, right?

913

:

That you might think, well, science is

boring, so we need to spice it up.

914

:

This happens all the time.

915

:

If you see scientists on daytime

television or whatever, they inevitably do

916

:

some chemistry experiment where there's

some explosion and gives people a really

917

:

distorted view of what science is.

918

:

Not only is it...

919

:

People think that it's old white dudes in

lab coats and geniuses, but also people

920

:

have this misconception that it's all

about excitement and explosions and

921

:

chemical reactions and cosmic awesomeness.

922

:

But science is at its core, this

fundamental framework for navigating the

923

:

world in the...

924

:

most sensible way possible.

925

:

So when I approach the children's books, I

try to really simplify not only the

926

:

concepts, but just that overall sense of

what I'm trying to do.

927

:

I'm not trying to create some extrapolated

vision, some way too exciting picture of

928

:

what science is.

929

:

What I try to do is I try to give

examples, analogies, categories, kind of

930

:

abstract things that give people some

comfort, some tools that they can use to

931

:

try to understand or appreciate what's

happening in these fields.

932

:

it becomes obvious that the books are for

parents, not necessarily for babies.

933

:

Um, and I think a lot of the feedback that

I get is from parents who say things like,

934

:

Oh, I wish I had learned this topic in

school in this way.

935

:

Right.

936

:

Uh, and you know, it all boils down to

this, this notion that when we learn

937

:

things, what, what we're doing is just

building up our repertoire of

938

:

of analogies that we can use to understand

them.

939

:

And the more that you have, the better,

right?

940

:

And the sooner you start, the better.

941

:

I think there is a misconception that

there's one unique special way to

942

:

understand a concept.

943

:

And if it's only told to you in that way,

some light bulb moment will happen in

944

:

which you all of a sudden understand it.

945

:

But that's just not

946

:

you at some point in the future, you say,

Oh, I feel like I understand that.

947

:

But there wasn't a, there wasn't a turning

point.

948

:

There wasn't a light bulb moment.

949

:

There wasn't a switch.

950

:

It was just time and, and building up

those, those analogies and examples that

951

:

at some point you just feel comfortable

and that's all there is to it.

952

:

So it's actually surprisingly easy.

953

:

It's a lot easier than people think.

954

:

Uh, you know, because the, the task that I

set myself is, is not such a high bar, you

955

:

know, just give a simple palatable analogy

for some core concept in the thing that

956

:

you're talking about that, that anyone can

understand.

957

:

Hmm.

958

:

Mm hmm.

959

:

Yeah.

960

:

Um, yeah, definitely.

961

:

It's.

962

:

Again, extremely important, so thanks a

lot for doing that.

963

:

And I do think that it's very important to

make science more, look more human and

964

:

write it more and more approachable

because I often people see that as very

965

:

dry endeavor, but I think actually

counting stories.

966

:

about science and scientists and normal

scientists, right?

967

:

Not the weird scientists from the movies

is extremely important because that's also

968

:

how we learn, right?

969

:

We learn a lot.

970

:

Our brain is like that.

971

:

We love stories and we love learning

through stories.

972

:

Like every equation you learned at school

has actually a story behind it.

973

:

Lots of people have worked on it.

974

:

Lots of people have.

975

:

failed and depressed because they couldn't

find the solution.

976

:

And thanks to their work, then afterwards

it unblocked a lot of things that you can

977

:

actually do now.

978

:

Just knowing about relativity makes us

able to be located through our phone.

979

:

We can use GPS very accurately because we

actually take into account relativity.

980

:

Well, it's pretty incredible, right?

981

:

I'm guessing most people don't know that.

982

:

So yeah, I think it's extremely important.

983

:

And actually I've watched very recently a

series, a Netflix series that does an

984

:

extremely good job, I found illustrating

science like that.

985

:

So it's still of course romanticized a

bit, but first the physics that's in the

986

:

show is pretty good and...

987

:

accurate, they don't refer to absolutely

completely crazy theories because the

988

:

series is called Lost in Space and the

beaches unite.

989

:

Something happened on Earth, I'm not going

to spoil it, but something happened on

990

:

Earth and then some people had to go and

try and colonize Alpha Centauri and we

991

:

follow the adventures of the families who

do that.

992

:

The science is pretty good on that and

also the depiction of the science is, I

993

:

found, very interesting.

994

:

We have some very interesting scenes where

it's like, oh, that's magic.

995

:

That's not magic.

996

:

That's math.

997

:

That was really cool.

998

:

I'm not going to spoil, but I definitely

recommend this series.

999

:

It's really well done.

:

01:10:22,422 --> 01:10:25,403

And of course, well, your book, Chris.

:

01:10:26,804 --> 01:10:32,306

And well, I think we could, we can call it

a show, I think, because I've already

:

01:10:32,306 --> 01:10:33,966

taken a lot of time from you.

:

01:10:33,966 --> 01:10:38,508

And for people watching the video, you can

see that the sun is setting down for me.

:

01:10:38,508 --> 01:10:42,050

So the, the luminosity is getting down.

:

01:10:42,350 --> 01:10:48,432

But I'd like, so before the last two

questions, my last question would be a bit

:

01:10:48,432 --> 01:10:49,933

of a general one.

:

01:10:50,093 --> 01:10:51,213

If you have any.

:

01:10:52,439 --> 01:10:58,708

advice, Chris, for students or young

researchers interested in quantum physics

:

01:10:58,708 --> 01:11:04,897

or even patient statistics, what advice

would you give them to start in these

:

01:11:04,897 --> 01:11:05,637

fields?

:

01:11:07,742 --> 01:11:14,845

Yeah, I think for young people that have

time on their hands, my advice is quite

:

01:11:14,845 --> 01:11:18,347

simple is to study mathematics.

:

01:11:18,708 --> 01:11:23,951

Mathematics is obviously the foundation of

statistics, also the foundation of quantum

:

01:11:23,951 --> 01:11:25,611

physics and all of physics.

:

01:11:25,832 --> 01:11:31,074

I see students coming into university who

are very excited about science.

:

01:11:31,074 --> 01:11:34,717

They come in, they say, I've read all of

Brian Greene's books and Stephen Hawking's

:

01:11:34,717 --> 01:11:35,477

books.

:

01:11:36,906 --> 01:11:38,506

I'm here to be a scientist.

:

01:11:39,646 --> 01:11:41,647

I live to be a quantum physicist.

:

01:11:41,787 --> 01:11:45,048

And then you hand them a test with only

math problems on it.

:

01:11:45,528 --> 01:11:49,709

And they get very deflated because nobody

told them that it was all about math.

:

01:11:50,569 --> 01:11:53,850

So it's the way that I came into the

field.

:

01:11:53,850 --> 01:11:56,731

I was never really interested in physics

or science.

:

01:11:56,731 --> 01:11:58,471

I was a math student.

:

01:11:58,591 --> 01:12:04,533

And when I finished my degree, it was more

about how am I going to apply my skills in

:

01:12:04,533 --> 01:12:06,253

solving math problems.

:

01:12:06,862 --> 01:12:09,022

And that served me very well.

:

01:12:09,062 --> 01:12:12,903

So yeah, become proficient at mathematics.

:

01:12:12,903 --> 01:12:16,524

There's lots of fun stuff in mathematics

when you, you know, at the surface level,

:

01:12:16,524 --> 01:12:18,465

depending on the way it's taught can feel

boring.

:

01:12:18,465 --> 01:12:24,587

And, but yeah, the further you dig deep

into it, the more interesting and more

:

01:12:24,587 --> 01:12:30,409

exciting it gets, and it will provide you

with a deeper understanding of the field

:

01:12:30,409 --> 01:12:33,469

that you end up applying it to then.

:

01:12:33,598 --> 01:12:38,101

than you could have ever imagined and

certainly more so than the people that are

:

01:12:38,101 --> 01:12:40,223

just still at that surface level.

:

01:12:40,583 --> 01:12:43,185

So yeah, that would be my advice.

:

01:12:43,226 --> 01:12:50,512

Also, especially for young people, for

students, life is very long and now is the

:

01:12:50,512 --> 01:12:54,715

time that you're encouraged to make

mistakes.

:

01:12:54,715 --> 01:13:01,341

And it's really the only time in your life

where you can make mistakes and get rapid

:

01:13:01,341 --> 01:13:02,361

feedback.

:

01:13:02,994 --> 01:13:06,474

And that's the thing that's encouraged and

that's the best way to learn.

:

01:13:06,494 --> 01:13:13,716

So, you know, approach it from that

perspective and also drag it out as long

:

01:13:13,716 --> 01:13:16,417

as you possibly can.

:

01:13:16,417 --> 01:13:17,077

Yeah.

:

01:13:18,098 --> 01:13:21,398

Completely agree with these

recommendations.

:

01:13:22,379 --> 01:13:31,421

Learn math and learn it well and take

risks very, very young and for the most

:

01:13:31,421 --> 01:13:32,521

time you can.

:

01:13:33,478 --> 01:13:36,800

Because yeah, that's definitely helpful.

:

01:13:37,080 --> 01:13:41,463

Even financially, like a good financial

advice, if you have to take risk and put

:

01:13:41,463 --> 01:13:44,985

all most of your money on stocks, that

would be when you're young and then when

:

01:13:44,985 --> 01:13:50,468

you get older, you get a bit less, a bit

more risk averse on your portfolio

:

01:13:50,468 --> 01:13:50,968

investment.

:

01:13:50,968 --> 01:13:54,710

Well, I would say that's the same thing

for life and for rapid feedback and

:

01:13:54,710 --> 01:14:00,053

failure when you are young and not having

your responsibilities to do that, you

:

01:14:00,053 --> 01:14:01,493

know, take the risks.

:

01:14:01,682 --> 01:14:02,822

And learn math.

:

01:14:03,283 --> 01:14:04,883

That's not a risk at all.

:

01:14:06,024 --> 01:14:06,685

Awesome, Chris.

:

01:14:06,685 --> 01:14:09,127

Well, I'm going to let you go.

:

01:14:09,127 --> 01:14:13,269

But before that, I'm going to ask you the

last two questions I gave a guest at the

:

01:14:13,269 --> 01:14:14,370

end of the show.

:

01:14:14,610 --> 01:14:19,413

First one, if you had unlimited time and

resources, which problem would you try to

:

01:14:19,413 --> 01:14:20,093

solve?

:

01:14:22,014 --> 01:14:25,995

I think that's easy, at least in my

discipline, I would build a large scale

:

01:14:25,995 --> 01:14:31,516

quantum computer and then I would set it

on the task of simulating various

:

01:14:31,516 --> 01:14:35,597

materials until it found a high

temperature or room temperature

:

01:14:35,597 --> 01:14:37,178

superconducting material.

:

01:14:37,478 --> 01:14:41,959

And then we'd build that and go, have free

energy around the world.

:

01:14:42,839 --> 01:14:44,500

That sounds nice.

:

01:14:44,660 --> 01:14:45,060

I love that.

:

01:14:45,060 --> 01:14:47,661

Yeah, awesome.

:

01:14:47,661 --> 01:14:50,201

You're the first one to answer that, but

love it.

:

01:14:51,378 --> 01:14:55,948

And second question, if you could have

dinner with any great scientific mind that

:

01:14:55,948 --> 01:14:58,613

alive or fictional, who would it be?

:

01:15:01,726 --> 01:15:06,987

Yeah, I mean, these sorts of questions I

think are difficult, especially for

:

01:15:07,688 --> 01:15:09,588

someone with an analytical brain.

:

01:15:09,988 --> 01:15:15,270

You know, you've got the one, the devil on

your shoulder saying, yeah, play along,

:

01:15:15,270 --> 01:15:16,870

it's a whimsical game.

:

01:15:19,611 --> 01:15:20,712

I thought about this actually.

:

01:15:20,712 --> 01:15:25,893

So I think there'd be some inherent

problems with obviously with a dead

:

01:15:25,893 --> 01:15:26,853

scientist.

:

01:15:27,782 --> 01:15:30,184

You know, there's obvious problems, but I

think the ones that people don't think

:

01:15:30,184 --> 01:15:37,471

about are Say, you know, I brought what I

Guess this is a magical scenario, but I

:

01:15:37,471 --> 01:15:40,714

don't know if it's I go back in time or

they come to our time But in some sense,

:

01:15:40,714 --> 01:15:45,318

it doesn't matter So I would prefer they

come to our time because you know, if go

:

01:15:45,318 --> 01:15:49,802

far enough in the past and they don't even

have toilets So let's bring them to our

:

01:15:49,802 --> 01:15:51,263

time, but there's a problem.

:

01:15:51,263 --> 01:15:54,270

Like if I brought Einstein here what

:

01:15:54,270 --> 01:15:55,150

what would I have to do?

:

01:15:55,150 --> 01:15:59,333

Would I have to explain a century of

advancements in like the actual field that

:

01:15:59,333 --> 01:16:00,074

he came up with?

:

01:16:00,074 --> 01:16:03,116

And would he even accept it?

:

01:16:03,116 --> 01:16:09,220

Like even in his lifetime, he refused to

accept all of the consequences of quantum

:

01:16:09,220 --> 01:16:09,620

physics.

:

01:16:09,620 --> 01:16:15,204

So, you know, it actually wouldn't be a

great conversation.

:

01:16:15,204 --> 01:16:20,608

I think scientists from the past would

just be, it would be too difficult to

:

01:16:20,608 --> 01:16:21,628

communicate.

:

01:16:23,382 --> 01:16:25,383

magically overcome say some language

barrier.

:

01:16:25,383 --> 01:16:30,667

Like they're, yeah, the contributions they

made obviously are timeless, but like that

:

01:16:30,667 --> 01:16:34,730

conversation that you could have wouldn't

be very insightful.

:

01:16:34,770 --> 01:16:38,713

So I feel like you'd have to go with a

living scientist, but then the problem

:

01:16:38,713 --> 01:16:43,176

with a living scientist is like, I can

just email them if I had a specific

:

01:16:43,176 --> 01:16:44,096

question.

:

01:16:44,757 --> 01:16:49,881

So it seems like far more, far easier

than...

:

01:16:50,978 --> 01:16:55,380

than organizing some dinner, which you can

have when you go to conferences anyway.

:

01:16:55,380 --> 01:16:59,222

So I've been to dinner with Nobel

laureates and stuff and celebrity

:

01:16:59,222 --> 01:17:02,964

scientists, and one of them was probably

enough.

:

01:17:04,005 --> 01:17:07,367

So then I think you're forced to go with a

fictional character.

:

01:17:07,367 --> 01:17:12,990

I don't know how many of your guests pick

a fictional character, but my favorite

:

01:17:12,990 --> 01:17:19,973

fictional character with a self-proclaimed

great mind is Marvin.

:

01:17:20,174 --> 01:17:23,114

paranoid android from the Hitchhiker's

Guide to the Galaxy.

:

01:17:23,114 --> 01:17:29,156

So I would uh, I'd have dinner with Marvin

and I know exactly what I'd ask him to.

:

01:17:29,156 --> 01:17:31,256

I'd ask him about AI alignment.

:

01:17:31,397 --> 01:17:37,318

Um, because I think it seems to, he seems

to have been solved with Marvin and uh, I

:

01:17:37,318 --> 01:17:44,100

think he would just give a wonderfully

nihilistic answer to what is AI alignment.

:

01:17:44,300 --> 01:17:45,700

Yeah.

:

01:17:45,700 --> 01:17:47,521

Yeah, no, that'd be fun.

:

01:17:47,521 --> 01:17:49,441

Yeah.

:

01:17:50,394 --> 01:17:51,534

I take part in this dinner.

:

01:17:51,534 --> 01:17:52,255

I don't know.

:

01:17:52,255 --> 01:17:53,715

Let me know when that happens.

:

01:17:55,236 --> 01:18:00,719

You want, oh, you want a bonus question,

uh, physics related, a choice like that.

:

01:18:00,719 --> 01:18:07,043

We had to make, uh, last time we did a

retreat at PIMC Labs, we do a retreat, uh,

:

01:18:07,043 --> 01:18:08,003

every year.

:

01:18:08,584 --> 01:18:11,365

And, uh, of course, it's just a bunch of

nerds getting together.

:

01:18:11,365 --> 01:18:14,987

So we always end up with, uh, very nerdy

questions.

:

01:18:15,407 --> 01:18:19,909

And, um, yeah, this year, I think one of

the main questions where

:

01:18:20,230 --> 01:18:28,777

So yeah, the year before, one of the main

questions was who would win in a plane

:

01:18:28,777 --> 01:18:34,921

war, so in an airplane war, Earth or

Jupiterians.

:

01:18:35,382 --> 01:18:42,007

And this year, but the one I want your

input on is this year was, if you could

:

01:18:42,007 --> 01:18:47,131

choose between these three options, which

one would you choose?

:

01:18:47,132 --> 01:18:49,738

If you could know what's...

:

01:18:49,738 --> 01:18:52,679

like what it's like to be in the quantum

realm?

:

01:18:53,961 --> 01:18:59,844

Or if you could go inside a black hole and

know what's there?

:

01:19:01,026 --> 01:19:08,031

Or if you could go to an alien planet and

meet them and talk with them, what would

:

01:19:08,031 --> 01:19:08,371

you choose?

:

01:19:08,371 --> 01:19:09,391

Right.

:

01:19:10,072 --> 01:19:12,153

Uh, there's only one, there's only one

correct choice.

:

01:19:12,153 --> 01:19:17,910

It's the third one because the other, the

other two, uh, would be bad.

:

01:19:17,910 --> 01:19:18,510

bad decisions.

:

01:19:18,510 --> 01:19:22,151

So it's the alien planet, yeah.

:

01:19:22,151 --> 01:19:23,331

There is no quantum realm.

:

01:19:23,331 --> 01:19:24,811

I wrote a blog post about that.

:

01:19:24,811 --> 01:19:27,052

I'll give you the link for the listeners.

:

01:19:27,952 --> 01:19:28,292

Oh, perfect.

:

01:19:28,292 --> 01:19:30,053

So you can't go there, obviously.

:

01:19:30,573 --> 01:19:37,095

There's technical challenges clearly with

shrinking a human, but also, yeah, our

:

01:19:37,095 --> 01:19:44,937

entire sense of perception is built on our

mesoscopic relationship with the world.

:

01:19:47,830 --> 01:19:51,452

Like clearly there'd be no sound, there'd

be no notion of sight.

:

01:19:53,134 --> 01:20:01,421

So even if you could get around this weird

idea of shrinking yourself, it wouldn't be

:

01:20:01,421 --> 01:20:03,442

a place to experience.

:

01:20:04,203 --> 01:20:08,026

And then inside a black hole, every

direction points down and you'd be

:

01:20:08,026 --> 01:20:08,666

spaghettified.

:

01:20:08,666 --> 01:20:09,347

So it's a bad idea.

:

01:20:09,347 --> 01:20:10,148

That'd be a problem.

:

01:20:10,148 --> 01:20:10,268

Yeah.

:

01:20:10,268 --> 01:20:12,590

I mean, I love that statement.

:

01:20:12,590 --> 01:20:14,070

So let's go to the alien planet.

:

01:20:16,032 --> 01:20:17,773

That's a technical term, actually.

:

01:20:18,546 --> 01:20:20,067

Yeah, yeah, yeah.

:

01:20:20,067 --> 01:20:22,948

Spaghettification.

:

01:20:22,948 --> 01:20:24,689

Yeah, yeah, yeah.

:

01:20:24,689 --> 01:20:29,632

And yeah, I mean, I'm shocked by the

revelation you just made on this podcast

:

01:20:29,632 --> 01:20:32,793

that Ant-Man is not a documentary.

:

01:20:33,514 --> 01:20:36,675

That's just, I'm just shocked.

:

01:20:36,675 --> 01:20:39,577

So I think it's time to stop the podcast.

:

01:20:40,037 --> 01:20:44,920

First of all, because I don't have any

more light and second, because well, I've

:

01:20:44,920 --> 01:20:46,581

taken a lot of time from you.

:

01:20:46,761 --> 01:20:48,061

Thanks a lot, Chris.

:

01:20:49,194 --> 01:20:50,555

That was really awesome.

:

01:20:51,656 --> 01:20:58,540

I learned a lot and we covered a lot of

topics so that was really perfect.

:

01:20:59,320 --> 01:21:03,643

As usual, I put resources and a link to

our website in the show notes for those

:

01:21:03,643 --> 01:21:04,964

who want to dig deeper.

:

01:21:05,104 --> 01:21:08,126

Thank you again, Chris, for taking the

time and being on this show.

More Episodes
99. #99 Exploring Quantum Physics with Bayesian Stats, with Chris Ferrie
01:07:31
98. #98 Fusing Statistical Physics, Machine Learning & Adaptive MCMC, with Marylou Gabrié
01:05:06
97. #97 Probably Overthinking Statistical Paradoxes, with Allen Downey
01:12:35
94. #94 Psychometrics Models & Choosing Priors, with Jonathan Templin
01:06:25
15. #15 The role of Python in Science and Education, with Michael Kennedy
01:05:52
29. #29 Model Assessment, Non-Parametric Models, And Much More, with Aki Vehtari
01:05:04
35. #35 The Past, Present & Future of BRMS, with Paul Bürkner
01:07:02
41. #41 Thinking Bayes, with Allen Downey
01:04:02
42. #42 How to Teach and Learn Bayesian Stats, with Mine Dogucu
01:05:59
45. #45 Biostats & Clinical Trial Design, with Frank Harrell
01:08:54
70. #70 Teaching Bayes for Biology & Biological Engineering, with Justin Bois
01:05:31