Joan Bruna Courant Institute and Center for Data Science New in ML A personal perspective from Academia
5yr PhD Startup in Applied experience between Mathematics MsC and PhD One Slide about 5yr Visiting myself experience as positions at Assistant Prof. UC industrial research Berkeley & NYU lab (Fair)
Foundations Mathematical of RL Foundations of DL One Slide about my ML group Geometric and Physics Deep Learning
Foundations Mathematical of RL Foundations of DL One Slide 10 PhDs 3 postdocs about my 1 MsC ML group Geometric 1 undergrad and Physics Deep Learning
You & Peers A PhD Journey Scientific Context Being a good ML citizen Calibrate expectations 2010s: DL Experimental and manage peer and gain independence Revolution pressure This talk
You & Peers A PhD Journey Scientific Context Being a good ML citizen Calibrate expectations 2010s: DL Experimental and manage peer and gain independence Revolution pressure This talk Aim Academic Perspective of ML career: a unique moment
Machine Computer translation Vision DL golden [Google, ’20] [Krizhevsky et al, 12-20] Games Science decade [AlphaGo, ’16] [AlphaStar, ’19] [AlphaFold, ’19]
Machine Computer translation Vision DL golden [Google, ’20] [Krizhevsky et al, 12-20] Games Science decade [AlphaGo, ’16] [AlphaStar, ’19] [AlphaFold, ’19]
Machine Computer translation Vision DL golden [Google, ’20] [Krizhevsky et al, 12-20] Games Science decade None of these problems was thought to be possible! [AlphaGo, ’16] [AlphaStar, ’19] [AlphaFold, ’19]
Role of Theory so far?
Role of Theory so Leo Breiman far? 1995
Role of Theory so Leo Breiman far? 1995 None of these questions is fully understood yet!
Role of Theory so Leo Breiman far? 1995 None of these questions is fully understood yet! We need YOUR help!
Can this go on? Data Hunger Compute Hunger
Critical Need for Theory Scaling-up approach unsustainable Can this go on? Data Hunger Compute Hunger
Role of ML Theory? ML (DL) Theory
Better Experiments Guiding principles Robustness Role of ML Guarantees Theory? ML (DL) Theory
Better Physical Experiments Sciences Guiding principles Scientific Computing Robustness Improved Sample Role of ML Guarantees Complexity Theory? ML (DL) Theory
Quantum Cosmology Mechanics [with B. Menard lab (JHU)] [Pfau et al. ’19] Parametrise wavefunctions with Build statistical models of early deep networks having right universe that explain expansion and symmetries non-Gaussianity Vignettes Distributional Robustness [Schmidt et al.’19]
Questions so far?
A (standard) PhD Journey Human knowledge [The illustrated guide to a Ph.D,.Matt Might]
A (standard) High- PhD school Journey Human knowledge [The illustrated guide to a Ph.D,.Matt Might]
Undergrad A (standard) High- PhD school Journey Human knowledge [The illustrated guide to a Ph.D,.Matt Might]
MsC Undergrad A (standard) High- PhD school Journey Human knowledge [The illustrated guide to a Ph.D,.Matt Might]
Catching up literature MsC Undergrad A (standard) High- PhD school Journey Human knowledge [The illustrated guide to a Ph.D,.Matt Might]
Catching up literature MsC Undergrad A (standard) High- PhD school Journey Human knowledge [The illustrated guide to a Ph.D,.Matt Might]
Catching up literature MsC Undergrad A (standard) PhD High- PhD school Journey Keep pushing! Human knowledge [The illustrated guide to a Ph.D,.Matt Might]
Some naive opinions Startup
No trajectory is better than others Some naive opinions Startup
No trajectory is better than others Some Non- markovian naive random opinions process Startup
No trajectory is better than others Some Non- markovian naive random opinions process Startup Outcome is important, but path too: human path
Year 0 The ML PhD Journey ML knowledge
Year 1 Kernels The ML PhD D Journey DL ML knowledge
Year 2 The ML IPM PhD Journey GAN ML knowledge
The ML PhD Journey ML field, ML field, year 0 year 5
Trends are often unpredictable The ML PhD Journey ML field, ML field, year 0 year 5
Trends are often unpredictable Dent The ML will push PhD knowledge— Journey wherever you land ML field, ML field, year 0 year 5
Trends are often unpredictable Dent The ML will push PhD knowledge— Journey wherever you land ML is remarkably broad: profit ML field, ML field, year 0 year 5 from it!
ML Academic Ecosystem
My initial view A few stars single- handedly pushing the field forward with breakthroughs ML Academic Ecosystem PhD Duties Follow my advisor Write a few papers
ML Academic Ecosystem
My current view ML research is primarily a team-effort. ML Academic Ecosystem
My current view ML research is primarily a team-effort. ML Academic Ecosystem Team can be: You & advisor Your lab-mates Github org etc.
Progress is My current view incremental ML research is primarily a team-effort. Good research should not discriminate. ML Academic Ecosystem Team can be: You & advisor Your lab-mates Github org etc.
Progress is My current view incremental ML research is primarily a team-effort. Good research should not discriminate. ML Academic Ecosystem Research is Team can be: essentially You & advisor sustained by Your lab-mates Github org students like etc. you !
Closing Personal Advice
Currently Less emphasis much emphasis on analyzing on designing new current methods algorithms, models, that work well in architectures. practice. Closing Personal Advice
Currently Less emphasis much emphasis on analyzing on designing new current methods algorithms, models, that work well in architectures. practice. Closing Personal Advice ML Civility Current ML Whenever A shared competition feels possible, shift focus responsibility daunting to from papers to (reviewing, everyone. ideas . teaching)
Currently Less emphasis much emphasis on analyzing on designing new current methods algorithms, models, that work well in architectures. practice. Closing Personal Advice ML Civility Current ML Whenever Feel A shared competition feels possible, shift focus free to responsibility daunting to from papers to rebuke! (reviewing, everyone. ideas . teaching)
Welcome to the field!
Recommend
More recommend