Data Science: The End of Statistics? Larry Wasserman Carnegie Mellon University Interface 2015
Data Science: The End of Statistics?
Larry WassermanCarnegie Mellon University
Interface 2015
Conclusion
Let’s turn the Interface meeting into the statistics version of NIPS
Conclusion
Let’s turn the Interface meeting into the statistics version of NIPS
This Talk
Will be short
Will be annoying provocative
This Talk
Will be short
Will be annoying provocative
This Talk
Will be short
Will be annoying provocative
Main Points
• Statisticians are being left out
• This should worry everyone (not just statisticians)
• It’s (partly) our fault
• We need a culture shift:
1. modernize training (no more UMVUE’s)
2. embrace the CS conference culture
3. watch and learn from CS: active learning, deep learning, SVM,online learning, RKHS, differential privacy ...
Main Points
• Statisticians are being left out
• This should worry everyone (not just statisticians)
• It’s (partly) our fault
• We need a culture shift:
1. modernize training (no more UMVUE’s)
2. embrace the CS conference culture
3. watch and learn from CS: active learning, deep learning, SVM,online learning, RKHS, differential privacy ...
Main Points
• Statisticians are being left out
• This should worry everyone (not just statisticians)
• It’s (partly) our fault
• We need a culture shift:
1. modernize training (no more UMVUE’s)
2. embrace the CS conference culture
3. watch and learn from CS: active learning, deep learning, SVM,online learning, RKHS, differential privacy ...
Main Points
• Statisticians are being left out
• This should worry everyone (not just statisticians)
• It’s (partly) our fault
• We need a culture shift:
1. modernize training (no more UMVUE’s)
2. embrace the CS conference culture
3. watch and learn from CS: active learning, deep learning, SVM,online learning, RKHS, differential privacy ...
Main Points
• Statisticians are being left out
• This should worry everyone (not just statisticians)
• It’s (partly) our fault
• We need a culture shift:
1. modernize training (no more UMVUE’s)
2. embrace the CS conference culture
3. watch and learn from CS: active learning, deep learning, SVM,online learning, RKHS, differential privacy ...
Where are the Statisticians?
• President’s Council of Advisors on Science and Technology(PCAST) includes ...
0 statisticians!
• Chief Data Scientist of the United States Office of Science andTechnology Policy.Not a statistician.
• Forbes: World’s 7 Most Powerful Data Scientists0 statisticians.
• Startups?• Google, Microsoft, Facebook all have Chief Economists. ChiefStatisticians?
Where are the Statisticians?
• President’s Council of Advisors on Science and Technology(PCAST) includes ...0 statisticians!
• Chief Data Scientist of the United States Office of Science andTechnology Policy.Not a statistician.
• Forbes: World’s 7 Most Powerful Data Scientists0 statisticians.
• Startups?• Google, Microsoft, Facebook all have Chief Economists. ChiefStatisticians?
Where are the Statisticians?
• President’s Council of Advisors on Science and Technology(PCAST) includes ...0 statisticians!
• Chief Data Scientist of the United States Office of Science andTechnology Policy.
Not a statistician.
• Forbes: World’s 7 Most Powerful Data Scientists0 statisticians.
• Startups?• Google, Microsoft, Facebook all have Chief Economists. ChiefStatisticians?
Where are the Statisticians?
• President’s Council of Advisors on Science and Technology(PCAST) includes ...0 statisticians!
• Chief Data Scientist of the United States Office of Science andTechnology Policy.Not a statistician.
• Forbes: World’s 7 Most Powerful Data Scientists0 statisticians.
• Startups?• Google, Microsoft, Facebook all have Chief Economists. ChiefStatisticians?
Where are the Statisticians?
• President’s Council of Advisors on Science and Technology(PCAST) includes ...0 statisticians!
• Chief Data Scientist of the United States Office of Science andTechnology Policy.Not a statistician.
• Forbes: World’s 7 Most Powerful Data Scientists
0 statisticians.
• Startups?• Google, Microsoft, Facebook all have Chief Economists. ChiefStatisticians?
Where are the Statisticians?
• President’s Council of Advisors on Science and Technology(PCAST) includes ...0 statisticians!
• Chief Data Scientist of the United States Office of Science andTechnology Policy.Not a statistician.
• Forbes: World’s 7 Most Powerful Data Scientists0 statisticians.
• Startups?• Google, Microsoft, Facebook all have Chief Economists. ChiefStatisticians?
Where are the Statisticians?
• President’s Council of Advisors on Science and Technology(PCAST) includes ...0 statisticians!
• Chief Data Scientist of the United States Office of Science andTechnology Policy.Not a statistician.
• Forbes: World’s 7 Most Powerful Data Scientists0 statisticians.
• Startups?
• Google, Microsoft, Facebook all have Chief Economists. ChiefStatisticians?
Where are the Statisticians?
• President’s Council of Advisors on Science and Technology(PCAST) includes ...0 statisticians!
• Chief Data Scientist of the United States Office of Science andTechnology Policy.Not a statistician.
• Forbes: World’s 7 Most Powerful Data Scientists0 statisticians.
• Startups?• Google, Microsoft, Facebook all have Chief Economists. ChiefStatisticians?
Everyone Should Care (Not Just Statisticians)
• Big Data + Bad Analysis = Bad Decisions
• Gary King: Big data is not about the data, it’s about theanalytics.
• Google search: big data bad analytics = 10,700,000 hits
• Statisticians have been doing data science for at least 100 years.
• You would not get brain surgery done by a cardiologist.
Everyone Should Care (Not Just Statisticians)
• Big Data + Bad Analysis = Bad Decisions
• Gary King: Big data is not about the data, it’s about theanalytics.
• Google search: big data bad analytics = 10,700,000 hits
• Statisticians have been doing data science for at least 100 years.
• You would not get brain surgery done by a cardiologist.
Everyone Should Care (Not Just Statisticians)
• Big Data + Bad Analysis = Bad Decisions
• Gary King: Big data is not about the data, it’s about theanalytics.
• Google search: big data bad analytics = 10,700,000 hits
• Statisticians have been doing data science for at least 100 years.
• You would not get brain surgery done by a cardiologist.
Everyone Should Care (Not Just Statisticians)
• Big Data + Bad Analysis = Bad Decisions
• Gary King: Big data is not about the data, it’s about theanalytics.
• Google search: big data bad analytics = 10,700,000 hits
• Statisticians have been doing data science for at least 100 years.
• You would not get brain surgery done by a cardiologist.
Everyone Should Care (Not Just Statisticians)
• Big Data + Bad Analysis = Bad Decisions
• Gary King: Big data is not about the data, it’s about theanalytics.
• Google search: big data bad analytics = 10,700,000 hits
• Statisticians have been doing data science for at least 100 years.
• You would not get brain surgery done by a cardiologist.
Why Are Statisticians Left Out?
Statisticians are:
conservativestubborninflexiblebad at selling themselvesafraidexperts at saying what you can’t do
Why Are Statisticians Left Out?
Statisticians are:
conservative
stubborninflexiblebad at selling themselvesafraidexperts at saying what you can’t do
Why Are Statisticians Left Out?
Statisticians are:
conservativestubborn
inflexiblebad at selling themselvesafraidexperts at saying what you can’t do
Why Are Statisticians Left Out?
Statisticians are:
conservativestubborninflexible
bad at selling themselvesafraidexperts at saying what you can’t do
Why Are Statisticians Left Out?
Statisticians are:
conservativestubborninflexiblebad at selling themselves
afraidexperts at saying what you can’t do
Why Are Statisticians Left Out?
Statisticians are:
conservativestubborninflexiblebad at selling themselvesafraid
experts at saying what you can’t do
Why Are Statisticians Left Out?
Statisticians are:
conservativestubborninflexiblebad at selling themselvesafraidexperts at saying what you can’t do
A (mostly) True Story
• Astronomer asks us for help.
• We spend months learning the science, cleaning the data andcarefully analyzing the data.
• Some careful, modest results after one year.
• In the meantime...... my astronomer friend went to see my friends in ML.
• Two days later the ML people produced fancy plots, analyses etc.
• We complain that their analysis was not rigorous.
• Who will the astronomer go to in the future?
A (mostly) True Story
• Astronomer asks us for help.
• We spend months learning the science, cleaning the data andcarefully analyzing the data.
• Some careful, modest results after one year.
• In the meantime...... my astronomer friend went to see my friends in ML.
• Two days later the ML people produced fancy plots, analyses etc.
• We complain that their analysis was not rigorous.
• Who will the astronomer go to in the future?
A (mostly) True Story
• Astronomer asks us for help.
• We spend months learning the science, cleaning the data andcarefully analyzing the data.
• Some careful, modest results after one year.
• In the meantime...... my astronomer friend went to see my friends in ML.
• Two days later the ML people produced fancy plots, analyses etc.
• We complain that their analysis was not rigorous.
• Who will the astronomer go to in the future?
A (mostly) True Story
• Astronomer asks us for help.
• We spend months learning the science, cleaning the data andcarefully analyzing the data.
• Some careful, modest results after one year.
• In the meantime...
... my astronomer friend went to see my friends in ML.
• Two days later the ML people produced fancy plots, analyses etc.
• We complain that their analysis was not rigorous.
• Who will the astronomer go to in the future?
A (mostly) True Story
• Astronomer asks us for help.
• We spend months learning the science, cleaning the data andcarefully analyzing the data.
• Some careful, modest results after one year.
• In the meantime...... my astronomer friend went to see my friends in ML.
• Two days later the ML people produced fancy plots, analyses etc.
• We complain that their analysis was not rigorous.
• Who will the astronomer go to in the future?
A (mostly) True Story
• Astronomer asks us for help.
• We spend months learning the science, cleaning the data andcarefully analyzing the data.
• Some careful, modest results after one year.
• In the meantime...... my astronomer friend went to see my friends in ML.
• Two days later the ML people produced fancy plots, analyses etc.
• We complain that their analysis was not rigorous.
• Who will the astronomer go to in the future?
A (mostly) True Story
• Astronomer asks us for help.
• We spend months learning the science, cleaning the data andcarefully analyzing the data.
• Some careful, modest results after one year.
• In the meantime...... my astronomer friend went to see my friends in ML.
• Two days later the ML people produced fancy plots, analyses etc.
• We complain that their analysis was not rigorous.
• Who will the astronomer go to in the future?
A (mostly) True Story
• Astronomer asks us for help.
• We spend months learning the science, cleaning the data andcarefully analyzing the data.
• Some careful, modest results after one year.
• In the meantime...... my astronomer friend went to see my friends in ML.
• Two days later the ML people produced fancy plots, analyses etc.
• We complain that their analysis was not rigorous.
• Who will the astronomer go to in the future?
Anecdote: My One Week as Editor of JASA
I was hired as editor of JASA.
I insisted that the journal be made freely available, online.
I was fired.
ASA sold the rights to the journal to Taylor and Francis.
JASA is still behind a paywall.
Compare this to JMLR (Journal of Machine Learning Research)jmlr.org. or NIPS (nips.cc) or ICML (imcl.cc) etc.
Anecdote: My One Week as Editor of JASA
I was hired as editor of JASA.
I insisted that the journal be made freely available, online.
I was fired.
ASA sold the rights to the journal to Taylor and Francis.
JASA is still behind a paywall.
Compare this to JMLR (Journal of Machine Learning Research)jmlr.org. or NIPS (nips.cc) or ICML (imcl.cc) etc.
Anecdote: My One Week as Editor of JASA
I was hired as editor of JASA.
I insisted that the journal be made freely available, online.
I was fired.
ASA sold the rights to the journal to Taylor and Francis.
JASA is still behind a paywall.
Compare this to JMLR (Journal of Machine Learning Research)jmlr.org. or NIPS (nips.cc) or ICML (imcl.cc) etc.
Anecdote: My One Week as Editor of JASA
I was hired as editor of JASA.
I insisted that the journal be made freely available, online.
I was fired.
ASA sold the rights to the journal to Taylor and Francis.
JASA is still behind a paywall.
Compare this to JMLR (Journal of Machine Learning Research)jmlr.org. or NIPS (nips.cc) or ICML (imcl.cc) etc.
Anecdote: My One Week as Editor of JASA
I was hired as editor of JASA.
I insisted that the journal be made freely available, online.
I was fired.
ASA sold the rights to the journal to Taylor and Francis.
JASA is still behind a paywall.
Compare this to JMLR (Journal of Machine Learning Research)jmlr.org. or NIPS (nips.cc) or ICML (imcl.cc) etc.
Anecdote: My One Week as Editor of JASA
I was hired as editor of JASA.
I insisted that the journal be made freely available, online.
I was fired.
ASA sold the rights to the journal to Taylor and Francis.
JASA is still behind a paywall.
Compare this to JMLR (Journal of Machine Learning Research)jmlr.org. or NIPS (nips.cc) or ICML (imcl.cc) etc.
Anecdote: My One Week as Editor of JASA
I was hired as editor of JASA.
I insisted that the journal be made freely available, online.
I was fired.
ASA sold the rights to the journal to Taylor and Francis.
JASA is still behind a paywall.
Compare this to JMLR (Journal of Machine Learning Research)jmlr.org. or NIPS (nips.cc) or ICML (imcl.cc) etc.
What to Do?
• Change “Department of Statistics” to “Department of Statisticsand Data Science”
• Mostly, we need a cultural shift: training, conferences, topics.
What to Do?
• Change “Department of Statistics” to “Department of Statisticsand Data Science”
• Mostly, we need a cultural shift: training, conferences, topics.
What to Do?
• Change “Department of Statistics” to “Department of Statisticsand Data Science”
• Mostly, we need a cultural shift: training, conferences, topics.
Training
• Get rid of: MVUE, ancillarity, completeness, ...
• Get rid of assumptions: (more on this is in a minute)
• Add:VC dimensionsupport vector machinesonline learning, banditsdeep learningoptimizationcoding (not just R)cloud computingbasic software engineering (github etc)
Training
• Get rid of: MVUE, ancillarity, completeness, ...
• Get rid of assumptions: (more on this is in a minute)
• Add:VC dimensionsupport vector machinesonline learning, banditsdeep learningoptimizationcoding (not just R)cloud computingbasic software engineering (github etc)
Training
• Get rid of: MVUE, ancillarity, completeness, ...
• Get rid of assumptions: (more on this is in a minute)
• Add:VC dimensionsupport vector machinesonline learning, banditsdeep learningoptimizationcoding (not just R)cloud computingbasic software engineering (github etc)
Training
• Get rid of: MVUE, ancillarity, completeness, ...
• Get rid of assumptions: (more on this is in a minute)
• Add:VC dimensionsupport vector machinesonline learning, banditsdeep learningoptimizationcoding (not just R)cloud computingbasic software engineering (github etc)
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Assumptions are For Suckers
• model-based, assumption-laden methods are useless in the worldof big, complex, datasets
• We need assumption-light methods with good visualization
• I propose we ban these things:
Y = Xβ + ε
Normality
sparsity (sparse methods not sparse models)
design assumptions (incoherence)
radical suggestion: let’s get rid of probability!
Jim Ramsay: “what good has probability ever done for statistics?”
Why should X1, . . . ,Xn be thought of as draws from somedistribution?
-online learning, individual sequence prediction, ...
Conference Culture
• conference model: refereed conferences: NIPS, ICML, AISTATS,etc• leads to energetic, fast, continuous progress• Every student should be regularly submitting papers to NIPS,AISTATS, ICML, ...• The Interface:-Let’s make the interface the epicenter of statistics. Make it likeNIPS.
Conference Culture
• conference model: refereed conferences: NIPS, ICML, AISTATS,etc
• leads to energetic, fast, continuous progress• Every student should be regularly submitting papers to NIPS,AISTATS, ICML, ...• The Interface:-Let’s make the interface the epicenter of statistics. Make it likeNIPS.
Conference Culture
• conference model: refereed conferences: NIPS, ICML, AISTATS,etc• leads to energetic, fast, continuous progress
• Every student should be regularly submitting papers to NIPS,AISTATS, ICML, ...• The Interface:-Let’s make the interface the epicenter of statistics. Make it likeNIPS.
Conference Culture
• conference model: refereed conferences: NIPS, ICML, AISTATS,etc• leads to energetic, fast, continuous progress• Every student should be regularly submitting papers to NIPS,AISTATS, ICML, ...
• The Interface:-Let’s make the interface the epicenter of statistics. Make it likeNIPS.
Conference Culture
• conference model: refereed conferences: NIPS, ICML, AISTATS,etc• leads to energetic, fast, continuous progress• Every student should be regularly submitting papers to NIPS,AISTATS, ICML, ...• The Interface:
-Let’s make the interface the epicenter of statistics. Make it likeNIPS.
Conference Culture
• conference model: refereed conferences: NIPS, ICML, AISTATS,etc• leads to energetic, fast, continuous progress• Every student should be regularly submitting papers to NIPS,AISTATS, ICML, ...• The Interface:-Let’s make the interface the epicenter of statistics. Make it likeNIPS.
Conclusion
• Statisticans are the original Data Scientists.
• Let’s embrace some of the CS culture. (If you can’t beat them,join them).
THE END
Conclusion
• Statisticans are the original Data Scientists.
• Let’s embrace some of the CS culture. (If you can’t beat them,join them).
THE END
Conclusion
• Statisticans are the original Data Scientists.
• Let’s embrace some of the CS culture. (If you can’t beat them,join them).
THE END
Conclusion
• Statisticans are the original Data Scientists.
• Let’s embrace some of the CS culture. (If you can’t beat them,join them).
THE END