learning biases in
play

Learning biases in opaque interactions BRANDON PRICKETT UNIVERSITY - PowerPoint PPT Presentation

Feb 10, 2024 •354 likes •1.3k views

Learning biases in opaque interactions BRANDON PRICKETT UNIVERSITY OF MASSACHUSETTS, AMHERST NECPHON 2017 Overview 1. Background 4. Conclusions 1. Phonological interactions 1. Which learner predicted the data best? 2. Opaque

  1. Learning biases in opaque interactions BRANDON PRICKETT UNIVERSITY OF MASSACHUSETTS, AMHERST NECPHON 2017

  2. Overview 1. Background 4. Conclusions 1. Phonological interactions 1. Which learner predicted the data best? 2. Opaque interactions 2. Future work 3. Kiparskian biases 4. Recent approaches to interaction learning 2. Experimental data 1. Methods 2. Results 3. Comparing the models’ performance 1. Expectation Driven Learning with SMR constraints 2. MaxEnt Stratal OT learner 3. A Sequence-to-Sequence neural net 2 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  3. 1. Background

  4. Phonological interactions • Phonological processes can interact with one another. • Sometimes these interactions are transparent in the surface form. 4 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  5. Phonological interactions • Phonological processes can interact with one another. • Sometimes these interactions are transparent in the surface form. • Example of a toy Feeding interaction (all examples adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kai/ /tai/ • Vowel Deletion: V  ∅ /_V -- ki ti • Palatalization: T  [+palatal]/_[+front] tʃi -- tʃi • SR [ tʃi ] [ki] [ tʃi ] 5 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  6. Phonological interactions • Phonological processes can interact with one another. • Sometimes these interactions are transparent in the surface form. • Example of a toy Feeding interaction (all examples adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kai/ /tai/ • Vowel Deletion: V  ∅ /_V Deletion moves the -- ki ti /i/ so that it can • Palatalization: T  [+palatal]/_[+front] tʃi -- tʃi trigger the /t/  [ tʃ ] • SR [ tʃi ] [ki] [ tʃi ] process. 6 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  7. Phonological interactions • Phonological processes can interact with one another. • Sometimes these interactions are transparent in the surface form. • Example of a toy Feeding interaction (all examples adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kai/ /tai/ • Vowel Deletion: V  ∅ /_V -- ki ti • Palatalization: T  [+palatal]/_[+front] tʃi -- tʃi • SR [ tʃi ] [ki] [ tʃi ] • Example of a toy Bleeding interaction: • UR /ti/ /kia/ /tia/ • Vowel Deletion: V  ∅ /_V -- ka ta • Palatalization: T  [+palatal]/_[+front] tʃi -- -- • SR [ tʃi ] [ka] [ta] 7 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  8. Phonological interactions • Phonological processes can interact with one another. • Sometimes these interactions are transparent in the surface form. • Example of a toy Feeding interaction (all examples adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kai/ /tai/ • Vowel Deletion: V  ∅ /_V -- ki ti • Palatalization: T  [+palatal]/_[+front] tʃi -- tʃi • SR [ tʃi ] [ki] [ tʃi ] • Example of a toy Bleeding interaction: • UR /ti/ /kia/ /tia/ • Vowel Deletion: V  ∅ /_V -- ka ta Deletion removes • Palatalization: T  [+palatal]/_[+front] the /i/ so that it tʃi -- -- can’t trigger the • SR [ tʃi ] [ka] [ta] /t/  [ tʃ ] process. 8 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  9. Opaque interactions • However, these interactions can also be opaque (Kiparsky 1968, 1971). • This can take the form of a process not applying when it seems like it should (counterfeeding) or applying when it seems like it shouldn’t (counterbleeding). • Example of a toy Counterfeeding interaction (adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kai/ /tai/ • Palatalization: T  [+palatal]/_[+front] tʃi -- -- • Vowel Deletion: V  ∅ /_V -- ki ti • SR [ tʃi ] [ki] [ti] 9 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  10. Opaque interactions • However, these interactions can also be opaque (Kiparsky 1968, 1971). • This can take the form of a process not applying when it seems like it should (counterfeeding) or applying when it seems like it shouldn’t (counterbleeding). • Example of a toy Counterfeeding interaction (adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kai/ /tai/ • Palatalization: T  [+palatal]/_[+front] tʃi -- -- Deletion moves the • Vowel Deletion: V  ∅ /_V /i/ next to the /t/, -- ki ti but does so too late • SR [ tʃi ] [ki] [ti] to cause the /t/  [ tʃ ] process. 10 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  11. Opaque interactions • However, these interactions can also be opaque (Kiparsky 1968, 1971). • This can take the form of a process not applying when it seems like it should (counterfeeding) or applying when it seems like it shouldn’t (counterbleeding). • Example of a toy Counterfeeding interaction (adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kai/ /tai/ • Palatalization: T  [+palatal]/_[+front] tʃi -- -- • Vowel Deletion: V  ∅ /_V -- ki ti • SR [ tʃi ] [ki] [ti] • Example of a toy Counterbleeding interaction (adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kia/ /tia/ • Palatalization: T  [+palatal]/_[+front] tʃi -- tʃia • Vowel Deletion: V  ∅ /_V -- ka tʃa • SR [ tʃi ] [ka] [ tʃa ] 11 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  12. Opaque interactions • However, these interactions can also be opaque (Kiparsky 1968, 1971). • This can take the form of a process not applying when it seems like it should (counterfeeding) or applying when it seems like it shouldn’t (counterbleeding). • Example of a toy Counterfeeding interaction (adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kai/ /tai/ • Palatalization: T  [+palatal]/_[+front] tʃi -- -- • Vowel Deletion: V  ∅ /_V -- ki ti • SR [ tʃi ] [ki] [ti] • Example of a toy Counterbleeding interaction (adapted from Baković 2011 and Jarosz 2016): • UR /ti/ /kia/ /tia/ Deletion removes the • Palatalization: T  [+palatal]/_[+front] /i/, but does so too tʃi -- tʃia • Vowel Deletion: V  ∅ /_V late to stop the -- ka tʃa /t/  [ tʃ ] process. • SR [ tʃi ] [ka] [ tʃa ] 12 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  13. Kiparskian biases • When discussing phonological interactions, Kiparsky (1968, 1971) discusses two different learning biases that could influence the likelihood of these interactions arising diachronically: • Application bias (Kiparsky 1968) – this favors interactions that cause more processes to apply. Out of the interactions we just discussed, this bias favors feeding and counterbleeding. • Transparency bias (Kiparsky 1971) – this favors interactions that apply processes in an order that is transparent from the surface form (i.e. there are no surface violations of rules that applied somewhere in the grammar). Out of the interactions we just discussed, this bias favors feeding and bleeding. • Kiparsky didn’t propose that both biases were at work, although I’ll be testing that possibility. 13 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  14. Kiparskian biases • When discussing phonological interactions, Kiparsky (1968, 1971) discusses two different learning biases that could influence the likelihood of these interactions arising diachronically: • Application bias (Kiparsky 1968) – this favors interactions that cause more processes to apply. Out of the interactions we just discussed, this bias favors feeding and counterbleeding. • Transparency bias (Kiparsky 1971) – this favors interactions that apply processes in an order that is transparent from the surface form (i.e. there are no surface violations of rules that applied somewhere in the grammar). Out of the interactions we just discussed, this bias favors feeding and bleeding. Application No Application Transparent Feeding Bleeding Opaque Counterbleeding Counterfeeding 14 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

  15. Kiparskian biases • When discussing phonological interactions, Kiparsky (1968, 1971) discusses two different learning biases that could influence the likelihood of these interactions arising diachronically: • Application bias (Kiparsky 1968) – this favors interactions that cause more processes to apply. Out of the interactions we just discussed, this bias favors feeding and counterbleeding. • Transparency bias (Kiparsky 1971) – this favors interactions that apply processes in an order that is transparent from the surface form (i.e. there are no surface violations of rules that applied somewhere in the grammar). Out of the interactions we just discussed, this bias favors feeding and bleeding. Very favored Application No Application Transparent Feeding Bleeding Opaque Counterbleeding Counterfeeding 15 University of Massachusetts, Amherst BPRICKETT@UMASS.EDU - HTTP://PEOPLE.UMASS.EDU/BPRICKETT/

Recommend

Using iterated learning to reveal biases for well-structured meanings in language Jon W. Carr

Using iterated learning to reveal biases for well-structured meanings in language Jon W. Carr

Using iterated learning to reveal biases for well-structured meanings in language Jon W. Carr Centre for Language Evolution University of Edinburgh Linguistics and English Language Postgraduate Conference 2016 What shapes language? Language

891 views • 62 slides
Capital Budgeting: Biases (Welch, Chapter 13-5) Ivo Welch More Biases Overconfidence Are you

Capital Budgeting: Biases (Welch, Chapter 13-5) Ivo Welch More Biases Overconfidence Are you

Capital Budgeting: Biases (Welch, Chapter 13-5) Ivo Welch More Biases Overconfidence Are you overconfident? (Note that this can also affect proper volatility estimates in real options.) 90% Confidence Interval Give a 90% confidence interval

517 views • 13 slides
Fairness and bias in Machine Learning A quick review on tools to detect biases in machine

Fairness and bias in Machine Learning A quick review on tools to detect biases in machine

QCon 2019 Fairness and bias in Machine Learning A quick review on tools to detect biases in machine learning model Thierry Silbermann, Tech Lead Data Science at Nubank thierry.silbermann@nubank.com.br Data collection Todays

870 views • 45 slides
On the Feasibility of Learning Human Biases for Reward Inference Rohin Shah, Noah Gundotra,

On the Feasibility of Learning Human Biases for Reward Inference Rohin Shah, Noah Gundotra,

On the Feasibility of Learning Human Biases for Reward Inference Rohin Shah, Noah Gundotra, Pieter Abbeel, Anca Dragan A conversation amongst IRL researchers A conversation amongst IRL researchers [Ziebart et al, 2008] To deal with suboptimal

487 views • 14 slides
Diversity & Inclusion Newsletter We All Have Implicit or Unconscious Biases Bias is defined

Diversity & Inclusion Newsletter We All Have Implicit or Unconscious Biases Bias is defined

Volume 1, Issue 2 Summer 2020 ASU Newport Diversity & Inclusion Newsletter We All Have Implicit or Unconscious Biases Bias is defined as prejudice in with the same experience and qualifications. D&I Learning Tools: favor of or

461 views • 3 slides
Natural Language Processing with Deep Learning Footprint of Societal Biases in NLP Navid

Natural Language Processing with Deep Learning Footprint of Societal Biases in NLP Navid

Natural Language Processing with Deep Learning Footprint of Societal Biases in NLP Navid Rekab-Saz navid.rekabsaz@jku.at Institute of Computational Perception Institute of Computational Perception Agenda Motivation Bias in word

822 views • 41 slides
Unconscious Bias 1 Questions to Start: Are we aware of our unconscious biases? Do we accept

Unconscious Bias 1 Questions to Start: Are we aware of our unconscious biases? Do we accept

Unconscious Bias 1 Questions to Start: Are we aware of our unconscious biases? Do we accept them? What types of biases are you aware of? Do you have experience with your own or seen biases in others? Activity: You are working

289 views • 14 slides
Breakout for small group discussions Topics Understanding your biases Awareness of

Breakout for small group discussions Topics Understanding your biases Awareness of

8/18/2020 Listen and Learn about Racial Justice 1 Presentation then Q&A Breakout for small group discussions Topics Understanding your biases Awareness of how biases affect behaviors Discerning your voice and your role

465 views • 13 slides
Investigating Potential Investigating Potential Biases in Aerosol Light Biases in Aerosol Light

Investigating Potential Investigating Potential Biases in Aerosol Light Biases in Aerosol Light

Investigating Potential Investigating Potential Biases in Aerosol Light Biases in Aerosol Light Absorption Absorption Measurements Measurements Christine Walsh Lund University (Lund, Sweden) NOAA ERSL (Aerosol Research Group) Overview of

354 views • 17 slides
Heuristics and biases Tina Nane 2 Heuristics and biases Lotto Icon by Dapete is

Heuristics and biases Tina Nane 2 Heuristics and biases Lotto Icon by Dapete is

1 Heuristics and biases Tina Nane 2 Heuristics and biases Lotto Icon by Dapete is licensed under CC BY 2.5 2 Heuristics and biases Heuristics Lotto Icon by Dapete is licensed under CC BY 2.5 Tricks, rules of thumb,

372 views • 23 slides
TEXT AND TEXT AND AUTOMATED BIASES AUTOMATED BIASES NATURAL LANGUAGES ARE THE NATURAL

TEXT AND TEXT AND AUTOMATED BIASES AUTOMATED BIASES NATURAL LANGUAGES ARE THE NATURAL

TEXT AND TEXT AND AUTOMATED BIASES AUTOMATED BIASES NATURAL LANGUAGES ARE THE NATURAL LANGUAGES ARE THE BASE HUMAN COMMUNICATION BASE HUMAN COMMUNICATION we learn from books of all kinds about complex topics and keep ourself updated

532 views • 51 slides
Biases in Decision Making Alexander Felfernig alexander.felfernig@ist.tugraz.at Decision Biases

Biases in Decision Making Alexander Felfernig alexander.felfernig@ist.tugraz.at Decision Biases

Institute for Software Technology International Workshop on Decision Making and Recommender Systems, Bolzano, 2014 Biases in Decision Making Alexander Felfernig alexander.felfernig@ist.tugraz.at Decision Biases & Recommender Systems 1

1.11k views • 53 slides
Spectral and Radiometric Issues for Level 1C L. Larrabee Strow and Scott Hannon Atmospheric

Spectral and Radiometric Issues for Level 1C L. Larrabee Strow and Scott Hannon Atmospheric

Introduction Scan Angle Biases Doppler Effect Biases IASI vs AIRS Biases Conclusions Spectral and Radiometric Issues for Level 1C L. Larrabee Strow and Scott Hannon Atmospheric Spectroscopy Laboratory (ASL) Physics Department and Joint

437 views • 20 slides
OVERCOMING RACISM IN SPITE OF OUR BLINDERS REV. ELIZABETH BIER & REV. JEN COLLINS SAINT PAUL

OVERCOMING RACISM IN SPITE OF OUR BLINDERS REV. ELIZABETH BIER & REV. JEN COLLINS SAINT PAUL

OVERCOMING RACISM IN SPITE OF OUR BLINDERS REV. ELIZABETH BIER & REV. JEN COLLINS SAINT PAUL AREA SYNOD OF ELCA LEARNING OBJECTIVES FOR TODAY To increase our awareness 1.) We all have biases 2.) To see how our biases are formed

299 views • 10 slides
Algorithmic Bias I: Biases and their Consequences Joshua A. Kroll Postdoctoral Research Scholar

Algorithmic Bias I: Biases and their Consequences Joshua A. Kroll Postdoctoral Research Scholar

Algorithmic Bias I: Biases and their Consequences Joshua A. Kroll Postdoctoral Research Scholar UC Berkeley School of Information 2019 ACM Africa Summer School on Machine Learning for Data Mining and Search 14-18 January 2019 Pe People

775 views • 62 slides
Optimization for Machine Learning Tom Schaul schaul@cims.nyu.edu Recap: Learning Machines

Optimization for Machine Learning Tom Schaul schaul@cims.nyu.edu Recap: Learning Machines

Optimization for Machine Learning Tom Schaul schaul@cims.nyu.edu Recap: Learning Machines Learning machines (Neural Networks, etc.) Forward passes produce a function of input Trainable parameters (aka weights, biases, etc.)

815 views • 50 slides
Ge Getting tting Re Real al Ab About out Bia Biases ses an and Issues d Issues Re

Ge Getting tting Re Real al Ab About out Bia Biases ses an and Issues d Issues Re

Ge Getting tting Re Real al Ab About out Bia Biases ses an and Issues d Issues Re Related lated to to Bi Bias as Kenne neth h Roach, ch, Ed.D .D., ., LCMHC HC Wendy dy Thomas, mas, MS, , ACMHC MHC The Universi versity

269 views • 26 slides
Bayesian Learning l A powerful approach in machine learning l Combine data seen so far with prior

Bayesian Learning l A powerful approach in machine learning l Combine data seen so far with prior

Bayesian Learning l A powerful approach in machine learning l Combine data seen so far with prior beliefs This is what has allowed us to do machine learning, have good inductive biases, overcome "No free lunch", and obtain good

1.42k views • 29 slides
Incentives and Behavior Prof. Dr. Heiner Schumacher KU Leuven 9. Behavioral Biases I Prof. Dr.

Incentives and Behavior Prof. Dr. Heiner Schumacher KU Leuven 9. Behavioral Biases I Prof. Dr.

Incentives and Behavior Prof. Dr. Heiner Schumacher KU Leuven 9. Behavioral Biases I Prof. Dr. Heiner Schumacher (KU Leuven) Incentives and Behavior 9. Behavioral Biases I 1 / 36 Introduction Literature Kahneman, Daniel (2011): Thinking,

774 views • 36 slides
Biases in NLP Models and What It Takes to Control them Kai-Wei Chang 1 A carton of ML (NLP)

Biases in NLP Models and What It Takes to Control them Kai-Wei Chang 1 A carton of ML (NLP)

Biases in NLP Models and What It Takes to Control them Kai-Wei Chang 1 A carton of ML (NLP) pipeline Prediction Evaluation (Structured) Inference Auxiliary Corpus/Models Representation (e.g, word embedding) Data Kai-Wei Chang

917 views • 73 slides
Perceptual Asymmetries in Learning Vowel Nasalization Kim Strtjen, Ruben van de Vijver, Dinah

Perceptual Asymmetries in Learning Vowel Nasalization Kim Strtjen, Ruben van de Vijver, Dinah

Perceptual Asymmetries in Learning Vowel Nasalization Kim Strtjen, Ruben van de Vijver, Dinah Baer-Henney OCP13 Budapest, 15.01.2016 Agenda 1 Learning biases 2 Vowel nasalization 3 Experiment 4 Results 5 Discussion 6 Conclusion 7 References

785 views • 39 slides
Errors, biases and Quality control in Next Gen Sequencing Dr David Humphreys

Errors, biases and Quality control in Next Gen Sequencing Dr David Humphreys

Errors, biases and Quality control in Next Gen Sequencing Dr David Humphreys d.humphreys@victorchang.edu.au - Lab scientist : Bioinformatician - RNA biologist - small RNAs (miRNA) Victor Chang Cardiac Research Institute, Sydney, Australia

423 views • 26 slides
Towards an Understanding of Human Persuasion and Biases in Argumentation Pierre Bisquert,

Towards an Understanding of Human Persuasion and Biases in Argumentation Pierre Bisquert,

Towards an Understanding of Human Persuasion and Biases in Argumentation Pierre Bisquert, Madalina Croitoru, Florence Dupin de Saint-Cyr, Abdelraouf Hecham INRA & LIRMM & IRIT, France July 6th 2016 B, C, D & H Persuasion and

531 views • 25 slides
BIASES IN THE ESTIMATION OF GALAXY CLUSTER DYNAMICAL MASSES 3.5m TNG 10.4m GTC mm Universe @

BIASES IN THE ESTIMATION OF GALAXY CLUSTER DYNAMICAL MASSES 3.5m TNG 10.4m GTC mm Universe @

BIASES IN THE ESTIMATION OF GALAXY CLUSTER DYNAMICAL MASSES 3.5m TNG 10.4m GTC mm Universe @ NIKA2, Grenoble 03-07 June 2019 Antonio Ferragamo THE CONTEXT PLANCK SZ CLUSTER CATALOGS PSZ1 PSZ2 Based on nominal mission data. Published in

441 views • 28 slides

More recommend