Successfully reported this slideshow.

Reverse inference problem

2

Share

Loading in …3
×
1 of 33
1 of 33

More Related Content

Reverse inference problem

  1. 1. Introduction: Revisiting reverse inference problem in functional MRI J. Chikazoe National Institute for Physiological Sciences
  2. 2. Annual Meeting of the Japan Neuroscience Society Disclosure of Conflict of Interest Name of first author: Junichi Chikazoe I have no COI with regard to the presentation.
  3. 3. Forward and reverse inference Forward inference Reverse inference Mental state induced by task (e.g. inhibition vs. control) P(A|M) A: Brain activity M: Mental state P(M|A) Mental state
  4. 4. Informal Reverse inference examples Conclusion: Participants may feel disgust because go/no-go task is too demanding. Response inhibition (task: go/no-go) The strongest activation in the insula Disgust (Wright et al., 2004)
  5. 5. Informal Reverse inference examples Conclusion: Participants may have to sustain task rule during the performance of go/no-go task. Response inhibition (task: go/no-go) The strongest activation in the insula Working memory (Engström et al., 2015)
  6. 6. Informal Reverse inference examples Conclusion: Participants may have ‘sweet’ feelings because successful responses in a difficult task can be taken as reward. Response inhibition (task: go/no-go) The strongest activation in the insula Gustation (Sweet) (Chikazoe et al., in preparation)
  7. 7. Informal Reverse inference examples Conclusion: Participants may have ‘sour’ feelings because… Response inhibition (task: go/no-go) The strongest activation in the insula Gustation (Sour) (Chikazoe et al., in preparation)
  8. 8. What is the issue of reverse inference? Bayes formula: P(M|A) = P(A|M) x P(M) + P(A|~M) x P(~M) P(A|M) x P(M) M: Mental process A: Brain activation
  9. 9. What is the issue of reverse inference? Bayes formula: P(M|A) = P(A|M) x P(M) + P(A|~M) x P(~M) P(A|M) x P(M) M: Mental process A: Activation estimated by arbitrarily picking up previous studies
  10. 10. What is the issue of reverse inference? Bayes formula: P(M|A) = P(A|M) x P(M) + P(A|~M) x P(~M) P(A|M) x P(M) M: Mental process A: Activation estimated by arbitrarily picking up previous studies estimated by our belief
  11. 11. Toward formal reverse inference What should we do?
  12. 12. Importance of pattern analysis for exploring shared neural correlates across modalities J. Chikazoe National Institute for Physiological Sciences
  13. 13. Cognitive components and their neural correlates Cognitive components Associated brain regions Decision making Posterior IPL Anterior MPFC PCC TPJ Self-referential processing Memory Multiple-to-multiple correspondence is observed.
  14. 14. Multiple functions in the same region -Most of cognitive functions may require multiple brain regions.  ( cf. connectionism )
  15. 15. Multiple functions in the same region -Most of cognitive functions may require multiple brain regions.  ( cf. connectionism ) →Global activation patterns may differ across functions.
  16. 16. Multiple functions in the same region -Most of cognitive functions may require multiple brain regions.  ( cf. connectionism ) →Global activation patterns may differ across functions. -The same region may have the similar computational processes but each neuron in that region may be assigned to different functions.
  17. 17. Multiple functions in the same region -Most of cognitive functions may require multiple brain regions.  ( cf. connectionism ) →Global activation patterns may differ across functions. -The same region may have the similar computational processes but each neuron in that region may be assigned to different functions. →Local activation patterns may differ across functions.
  18. 18. Multiple functions in the same region -Most of cognitive functions may require multiple brain regions.  ( cf. connectionism ) →Global activation patterns may differ across functions. -The same region may have the similar computational processes but each neuron in that region may be assigned to different functions. →Local activation patterns may differ across functions. -The same neural correlates may be shared across functions.
  19. 19. Multiple functions in the same region -Most of cognitive functions may require multiple brain regions.  ( cf. connectionism ) →Global activation patterns may differ across functions. -The same region may have the similar computational processes but each neuron in that region may be assigned to different functions. →Local activation patterns may differ across functions. -The same neural correlates may be shared across functions.
  20. 20. Positivity and negativity Kringelbach and Rolls 2004 A meta-analysis demonstrated that positive value is represented in the medial OFC, while negative value is represented in the lateral OFC.
  21. 21. Contradicting evidence from monkey electrophysiological studies Positivity- and negativity-sensitive neurons are interspersed in the OFC. (Morrison et al., 2009)
  22. 22. Averaged brain activity does not have sufficient specificity FMRI data showed overlap between the positivity- and negativity-sensitive regions. (Chikazoe et al., 2014)OverlapPositive Negative
  23. 23. Global or local activation patterns Global activation patterns Positivity Negativity
  24. 24. Global or local activation patterns Global activation patterns Local activation patterns Positivity Negativity
  25. 25. Neurosynth (created by Dr. Yarkoni) Meta-analysis P(pain|activation) Automated coordinate extraction Related studiesTerm-based search ‘Pain’ Yarkoni et al., 2011, Nature Methods P(M|A) = P(A|M) x P(M) + P(A|~M) x P(~M) P(A|M) x P(M) estimated, based on almost all fMRI studies set to 0.5 (uninformative prior)
  26. 26. Global activation patterns Positivity-related activity Positive Negativity-related activity Negative 0.13 0.07
  27. 27. Global activation patterns Positivity-related activity Positive Negativity-related activity Negative 0.13 0.070.05 0.08
  28. 28. Representational similarity analysis Local activation pattern analysis revealed that positivity and negativity could be discriminated. vectorize Creating RSM Local activation patterns Trial k Trial l Trial-by-trial correlation Value representational similarity matrix Value PosNeg Neu Neg Pos Neu Value SimilarDissimilar
  29. 29. Bayesian regression analysis H0 : β3 = 0 vs. H1 : β3 > 0 (H0 corresponds to ‘no relationship between neural and valence representations’) (Chikazoe et al., 2014)
  30. 30. Bayes factor estimation Representational similarity matrix BF10 (univariate) BF10 (pattern) Visual X Visual Visual X Gustatory Gustatory X Gustatory <.01 <.01 <.01 19 (Decisive for null) (Decisive for null) (Decisive for null) (Strong for H1) (Strong for H1)(Strong for H1) 39>100 Visual value Gustatory value Gustatory value Visualvalue Visualvalue Pos PosNeg Neg Neu Neu Gustatoryvalue Pos Neg Neu Pos PosNeg Neg Neu Neu PosNeg Neu
  31. 31. Summary -Global and local activation patterns were useful for formal reverse inference. -Shared neural correlates should satisfy cross-condition correspondence as well as within-condition correspondence.
  32. 32. Acknowledgements Cornell University Dr. Adam Anderson Dr. Eve de Rosa Ross Makello Columbia University Dr. Nikolaus Kriegeskorte National Institute for Physiological Sciences Dr. Norihiro Sadato Takaaki Yoshimoto Dr. Balbir Awana Ryutaro Uchiyama Funded by the Imaging Science Project of the Center for Novel Science Initiatives (CNSI)(# IS281004)
  33. 33. Decoding as a special case of reverse inference Bayes formula: P(M|A) = P(A|M∩Task) x P(M|Task) + P(A|~M ∩Task) x P(~M|Task) P(A|M∩Task) x P(M|Task) M: Mental process A: Activation pattern (Hutzler 2014) For “decoding” or local activation pattern analysis, we do (can) not consider other tasks (experiments).

Editor's Notes

  • Before moving to each presentation, I would like to briefly explain the purpose of this symposium.
  • I will briefly explain what forward and reverse inference are.
    In forward inference, probability of brain activation given mental state is inferred.
    This is the basic form of fMRI studies.
    In reverse inference, mental state is inferred from brain activation.
  • Let’s say we conducted an fmri study investigating brain regions associated with response inhibition.
    In our dataset, we found the strongest activation in the insula.
    We can say the insula is associated with response inhibition.
    This is forward inference.
    From this activation, we may want to infer reversely.
    For example, the insula is known to be associated with disgust.
    By applying reverse inference, we may draw a conclusion
    That participants may feel disgust because response inhibition task is too demanding.
  • However, another study demonstrated strong relationship between the insula and working memory.
    So, in this case our conclusion will be that participants may have to sustain task rule during the performance of go/no-go task.
  • Another study provided another evidence.
    The insula is known as the primary gustatory cortex.
    Sweet liquid evoked insula activation.
    Based on this result we may draw a conclusion that participants may have sweet feelings because successful responses in a difficult task can be taken as reward.
  • But the same study demonstrated sour taste activates the insula.
    So, we may have to draw a different conclusion from this.
  • Such a situation may look like a comedy, but what is the issue of the reverse inference?
    In reverse inference, we estimate the conditional probability of a mental process given a brain activation result.
  • In the informal reverse inference, we estimate conditional probability of brain activation given a mental state by arbitrarily picking up previous studies.
  • Furthremore, prior probability is estimated by our own belief.
    So, the conclusion is strongly biased.
  • In this symposium, we will discuss how we can apply reverse inference formally.
  • First, I would like to talk about relationship between cognitive components and their neural correlates.
    In most cases, 1-to-1 correspondence is not observed.
    For example, self-referential processing is associated with anterior MPFC as well as PCC and TPJ.
    Conversely, the anterior MPFC is related with self-referential processing, decision making and memory.
    In this way, multiple-to-multiple correspondence is observed.
  • This suggests multiple functions in the same region.
    For example, most of cognitive functions may require multiple brain regions.
    In this case, global activation patterns may differ across functions.
    Another possibility is that the same region may have the similar computational processes but each neuron in that region may be assigned to different functions.
    In this case local activation patterns may differ across functions.
    Another possibility is that the same construct may
  • This suggests multiple functions in the same region.
    For example, most of cognitive functions may require multiple brain regions.
    In this case, global activation patterns may differ across functions.
    Another possibility is that the same region may have the similar computational processes but each neuron in that region may be assigned to different functions.
    In this case local activation patterns may differ across functions.
    Another possibility is that the same construct may
  • This suggests multiple functions in the same region.
    For example, most of cognitive functions may require multiple brain regions.
    In this case, global activation patterns may differ across functions.
    Another possibility is that the same region may have the similar computational processes but each neuron in that region may be assigned to different functions.
    In this case local activation patterns may differ across functions.
    Another possibility is that the same construct may
  • This suggests multiple functions in the same region.
    For example, most of cognitive functions may require multiple brain regions.
    In this case, global activation patterns may differ across functions.
    Another possibility is that the same region may have the similar computational processes but each neuron in that region may be assigned to different functions.
    In this case local activation patterns may differ across functions.
    Another possibility is that the same construct may
  • This suggests multiple functions in the same region.
    For example, most of cognitive functions may require multiple brain regions.
    In this case, global activation patterns may differ across functions.
    Another possibility is that the same region may have the similar computational processes but each neuron in that region may be assigned to different functions.
    In this case local activation patterns may differ across functions.
    Another possibility is that the same construct may
  • This suggests multiple functions in the same region.
    For example, most of cognitive functions may require multiple brain regions.
    In this case, global activation patterns may differ across functions.
    Another possibility is that the same region may have the similar computational processes but each neuron in that region may be assigned to different functions.
    In this case local activation patterns may differ across functions.
    Another possibility is that the same construct may
  • From here, I would like to talk about my study.
    I am interested in value representations.
  • Our fmri study showed consistent results.
    In this study, emotionally positive or negative stimuli were presented, and participants rated positivity and negativity for each stimulus.
    This slide shows univariate analysis results.
    Yellow indicates brain regions sensitive to positive stimuli, and blue indicates negativity-sensitive regions.
    Green indicates overlap.
    This shows large overlap between positive and negative regions.
  • How can we discriminate such activation patterns?
    One possible solution will be employing multivoxel pattern analysis.
    For that purpose, we can use global activation patterns or local activation patterns.
  • How can we discriminate such activation patterns?
    One possible solution will be employing multivoxel pattern analysis.
    For that purpose, we can use global activation patterns or local activation patterns.
  • Neurosynth created by Dr. Yarkoni is a very powerful tool to perform reverse inference using global activation patterns.
    This online software automatically searches published fMRI studies and stores information of text and activation coordinate.
    Based on a huge number of published papers, probability of activation given a mental state is estimated.
    This prior probability can be calculated from the empirical published data, but for the purpose of comparison,
    This is set to 0.5.
  • Using Neursynth, I’ve got correlation between my imaging result and reverse inference map created by Neurosynth.
    This shows positivity-related activity is more similar to the reverse inference map associated with the term positive
    while
  • Negativity-related activity map is similar to the reverse inference map related to the term negative.
  • Another direction is performing MVPA using local activation patterns.
    I extracted the brain activity data from medial OFC and then vectorized them.
    Correlations between those vectors are calculated, resulting in representational similarity matrix.
    The combination of negative and negative or positive and positive shows higher correlation,
    While the combination of positive and negative shows lower correlation.
    This indicates
  • In the previous paper, we decomposed neural representational similarity in the OFC into several components such as visual feature, categories and valence using multiple regression analysis.
    This time, we applied bayesian regression analysis on this equation.
    The null hypothesis is no relationship between neural and valence representations.
  • We calculated bayes factor for 3 matrices.
    The first one is within-visual comparison, the second one is within gustatory comparison,
    And the third one is cross-modal comparison.
    We compared bayes factor obtained by univariate and pattern analysis.
    While univariate analysis failed to show strong association between neural and valence representations in the OFC,
    Multivariate analysis showed strong association between them.
    Importantly, not only within modal comparison, but also cross modal comparison shows strong association between neural and valence representations in the OFC.
  • When analyzing local activation patterns, we cannot compare them to other studies.
    This means all the terms in the Bayes formula were conditioned by task.
  • ×