[ERP Decoding] The step toward mind reading
2023. 2. 10. 00:15
๐ง Key Words
fMRI
ERP scalp map
orientations
์ด๋ป๊ฒ ์ฌ๋์ ๋ง์์ ์ฝ์ ๊ฒ์ธ๊ฐ?
- ์ฒ์์ผ๋ก ์ ๊ทผํ ๊ฒ์ mind reading์ ์ํด visual orientation์ ์ดํผ๋ ๊ฒ์ด๋ค.
- Different orientations maximally stimulate different fMRI voxels in visual cortex.
- ๋ค๋ฅธ orientation ์๊ทน์ ๋ํด fMRI๋ V1 ์์ญ์์ ๋ค๋ฅธ ๋ถ๋ถ์ ๋ํ๋ธ๋ค.
- We might therefore expect slightly different ERP scalp distributions(i.e., voltage pattern over the scalp) for different orientations.
- ๊ทธ๋ฌ๋ฏ๋ก ์ฐ๋ฆฌ๋ ๋ค๋ฅธ orientation ์๊ทน์ ๋ํด ERP๋ ๋ถ์์์๋ ERP scalp map์ด ๋๋ฅด๊ฒ ๋ํ๋๋ ๊ฒ์ ๊ธฐ๋ํ ์ ์๋ค.
- ์คํ๋ด์ฉ.
- 16๊ฐ์ orientations, Scalp distribution of ERP for 16 orientations.
- EEG recording.
- ํด๋น orientation์ ์ฃผ๊ณ ์ ์ ๋ค ๊ทธ orientation์ ํผ์คํ์๊ฐ ์ฐ๋๋ก ํ๋ ๊ฒ.
- next
- ERP scalp distrubution with orientation label 16๊ฐ์ ๋ํด ์๋ก ๊ตฌํ dataset์ธ scalp map์ ๋น๊ตํ์ฌ ์ด๋ค scalp map์ด ์๋ก ๊ตฌํ ์ด data์ ๊ฐ์ฅ ๋น์ทํ๊ฐ๋ฅผ ๋ณธ๋ค.
- ์ฐพ์ ๊ฒ์ label์ orientation์ด 60๋ ์ด๋ฉด ์๋ก ๊ตฌํ data set์ orientation๋ 60๋ ์ผ ๊ฒ์ด๋ค.
- If the scalp distributions for different orientations are distinguishable,
- and if the differences in the scalp distributions are systematic,
- then we should be able to guess the orientation associated with this new scalp distributon.
Decoding orientation from the scalp distribution of ERP
- y์ถ์ Decoding Accuracy, x์ถ์ chance level๋ก time.
- x์ถ์์,
- Stimulus Encoding
- Working memory maintenance
- Orientation report
- ์ ๊ฐ์ ์์๋ก ์ด๋ฃจ์ด์ง ์ ์๋ค.
- If the scalp ERP contains no information about the orientation,
- then decoding accuracy should be at chance.
- ERP scalp๊ฐ orientation ์ ๋ณด๋ฅผ ํฌํจํ๋ฉด chance ์ด์์ด๊ณ , ์๋๋ผ๋ฉด decoding accuracy๊ฐ chance level์ผ ๊ฒ์ด๋ค.
- Gray area : Statistically significant time points.
- ๊ทธ๋ฆฌ๊ณ ์๊ทน์ ๋ณด๋ perceptual ๊ธฐ๊ฐ๊ณผ ์๊ทน์ ๊ธฐ์ตํ๋ working memory ๊ธฐ๊ฐ์ผ๋ก ๋๋๋ค.
- ์ gray area์๋ working memory ๊ธฐ๊ฐ๋ง ํต๊ณ์ ์ธ ๋ถ์์ ํ๋ค.
- ๋ํ๋ฅผ ํตํด ์ด๋ค orientation์ ๊ธฐ์ตํ๊ณ ์๋์ง Decoding ํ ์ ์๋ค.
- Spatial distribution of ERP contains information about orientation !
- orientation์ด ๊ธฐ์ต๋๊ณ ์์์ ๋ณด์ฌ์ค๋ค.
์ ๋ง Orientation ๋ณ๋ก Decoding์ด ๋๋ ๊ฑธ๊น?
Confusion Matrix
- stimulus, classification, probability๊ฐ ์กด์ฌํ๋ค.
- ๋๋ถ๋ถ์ orientation๋ค์ด ๋ค Decoding ๋จ์ ๋ณด์ฌ์ค๋ค.
- Decoding was above chance for most of the orientations !
demonstration ์ด ์๋ฃ ๋์๋ค.
์ด๋ฅผ ํตํด ์ฐ๋ฆฌ๊ฐ ๋ญ ํ ์ ์์๊น?
Experiment 2.
- ์ด๋ฒ์๋ ๊ธฐ์ตํด์ผํ๋ ์๊ทน์ด ๋์ค๋ ์์น์ ๋ฐ์ํด์ผํ๋ ๊ฒ์ด ๋์ค๋ ์์น๊ฐ random์ผ๋ก ์ค์ ํ์๋ค.
- Attention์ ๋ณด๊ธฐ ์ํจ์ด๋ค.
- Location of sample teardrop and test teardrop varied randomly every trial.
- Attending to this location is not a useful strategy.
- Orientation value์ Location value๋ฅผ ์์ ํ randomํ๊ฒ independentํ๊ฒ ์ค์ํ๋ค.
- ์ด์ trial์ ๋ณธ ์๊ทน์ ํ์ฌ trial์ decodingํ ์ ์๋.
- Information of the previous-trial orientation mush be present during the current trial.
- The two orientations were independent.
- But the report of the current trial was biased by the orientation from the previous trial.
- Decoding ๋๋ค๊ณ ํ๋ฉด, ๊ณผ๊ฑฐ ์ํ์ด ํ์ฌ์๋ ๋จ์ ์๋ค.
- ์ด์ ์ํ์ ๋ดค๋ ๊ฒ์ด ํ์ฌ ์ํ์์๋ ์ํฅ์ ๋ผ์น๋ ๊ฒ์ ์ ์ ์๋ค.
- Memory reactivation์ Decoding์ ํตํด ์ ์ ์๋ค.
- motion direction ๊ณผ ๊ฐ์ ๊ฒ์์๋ ๊ฐ๋ฅ. ( Continuous direction estimation task )
- ์ ํด์ง ์๊ฐ์ motion direction ์ ๋ณด๋ฅผ ์ถ์ ํ์ฌ ๋ฐฉํฅ์ ๋ง์ถ์ด์ผ ํจ.
- ERP-based stimulus decoding:
- Spatial pattern of sustained ERP activity reflects both direction information and shift of attention.
- Decoding faces from the scalp distribution of ERP.
- ERP Decoding์์ Indentity or facial expression ๋ ๋ค decoding ๊ฐ๋ฅํ๊ฐ.
- We can decode ID independently of expression and expression independently of ID.
- Decoding Positive vs. Negative images
- ๊ธ์ , ๋ถ์ ์ฌ์ง์ ๋ํด ์ด๋ ํ ERP ์ฐจ์ด๋ฅผ ๋ณด์ผ ๊ฒ์ธ๊ฐ.
- ๊ฐ ์์ฒด๋ ์ฐจ์ด๋ฅผ ๋ณด์ด์ง ์์ง๋ง, Decoding accuracy๋ ๊ทน๋ช ํ ์ฐจ์ด๋ฅผ ๋ณด์ธ๋ค.
- Barely any difference between positive and negative images.
- Why is decoding more sensitive than univariate ERP analysis?
- univariate ์์ค์์ ํ๋จํ๊ธฐ ์ด๋ ค์ด ์์ค์ ์ฐจ์ด๋ฅผ multivariate analysis์์ ๋ ์ ๊ตฌ๋ถํ ์ ์๋ค.
Limitaions in the ERP decoding analysis
- Decoding is done from averaged ERPs
- ํ๊ท ์ ๋ด์ ํ๋จํ๋ฏ๋ก ๊ณตํ์ ์ผ๋ก ํ๊ณ๊ฐ ์กด์ฌํจ.
- Single-trial decdoing is not reliable.
- However, depending on the signal, reliable decoding may be possible.
- Decoding accuracy is far from 100%.
- Decoding accuracy๊ฐ ๋ฎ์ ๊ฒ์ real world์์๋ ํ๊ณ์ด์ง๋ง scientificํ ๊ฒ์์๋ ์ด์ฉ๋ ์ ์๋ค.
- Suitable for answering scientific questions.
- Not suitable for engineering applications.
- Usually, we do not see reliable correlation between decoding accuracy and behavioral performance.
- decoding์ด ์ข์ผ๋ฉด ํ๋ ๋ฅ๋ ฅ์ด ์ข์ ๊ฒ์ธ๊ฐ?
- Individual differences in decoding accuracy are probably driven mainly by cortical geometry and signal-to-noise ratio.
- ์ฌ๋๋ง๋ค ๋ค๋ฅธ ๊ฒ ๋๋ฌธ์ decoding์ ์ฐจ์ด๊ฐ ์์ ๋ฟ, ํ๋ ๋ฅ๋ ฅ๊ณผ๋ ๊ด๋ จ ์๋ค.
- We do not know which neural process is being used in the decoding.
- ์ด๋ค Neural process๊ฐ decoding์ ์ฐจ์ด๋ฅผ ๋ณด์ด๋์ง ์ ์๊ฐ ์๋ค.
- ๋ ๊ฐ ์ฌ์ด์ ์ฐจ์ด๋ง ์์ผ๋ฉด classify๊ฐ ๊ฐ๋ฅํ๊ณ ์ด๋ฅผ ํตํด above chance decoding์ด ๊ฐ๋ฅํ๋ค.
- ๊ทธ๋ฌ๋ ์ด๋ค ์ ํธ๊ฐ ๊ทธ๊ฒ์ ๋ง๋ค์ด ๋ด๋์ง๋ ์ ์๊ฐ ์๋ค.
ERP Analysis์ ERP Decoding ๋ ๊ฐ๋ ์๋ก์ ์ฒ๋ฆฌ์ ๊ดํด ์์ญ์ด ๋ค๋ฅผ ๋ฟ, ๋ ๋ค ์๋ฏธ ์๋ ๋ถ์ ๋ฐฉ๋ฒ์ด๋ค.
How is decoding different from the standard approach ?
- We can separate signals associated with the representational contents from signals associated with contents-free support process.
- ERP๋ ํ๋ก์ธ์ค๋ฅผ ๋ณด๋ ๊ฒ์ผ๋ก, ํ๋ก์ธ์ค๊ฐ ERP Amplitude๊ฐ ์ด๋ค ๊ฒ์ธ์ง ์ ์ ์๋ค.
- Decoding๋ ์ด์ ๋ค๋ฅด๊ฒ content์ ๋ํ, ๋ญ๊ฐ ์ง๊ฐ๋๋์ง๋ฅผ ์ ์ ์๋ค.
- Decoding answers the quesion "Does t he neural signal contain information about the stimulus?" with minimal assumptions about how the information is encoded in the neural signal
- We do not know which neural process is being used in the decoding
- ์ด ERP pattern์ด ์ด๋ค ์ ๋ณด๋ฅผ ๊ฐ์ง๊ณ ์๋๋. ์ฐจ์ด๊ฐ ์กด์ฌํ๊ธฐ๋ง ํ๋ค๋ฉด ๋ถ๋ฆฌํ์ฌ ๋ถ์ ๊ฐ๋ฅํ๋ค. Assumption์ด ์๋ ๊ฒ์ด ์ฅ์ .
- ๊ทธ๋ฌ๋ ์ด๋ค specificํ neural process๊ฐ decoding above chance๋ฅผ ๋ง๋ค์ด๋ด๋์ง๋ ์ ์ ์๋ค.
- Decoding is done separately for each subject, whereas the standard approach looks for consistency across subjects.
- Greater statistical power in basic science research.
- Decoding์ subject ๋ง๋ค ๋ฐ๋ก ์ด๋ฃจ์ด์ง๋ค.
- Individual difference๊ฐ ์ํฅ์ด ์์ง๋ง, ERP์์๋ ์ค์ํ๋ค.
- Decoding์ ๊ฐ๊ฐ ๋ถ์์ด ๋๊ธฐ ๋๋ฌธ์ individual difference๊ฐ ์๋ค๊ณ ํ๋๋ผ๋ ํจ๊ณผ๊ฐ ์ ๊ฒ ๋๋ค.
- ์ด ์ฐจ์ด๋ฅผ ๋ณด๊ธฐ ์ํด Effect Size๋ฅผ ๋ณธ๋ค.
๋ง์ฝ Decoding์ด above chance๊ฐ ์๋๋ผ๋ฉด ?
- If decoding accuracy of one condition/group is worse than another, this may be because the data are noisier in that group/condition.
- Decoding accuracy depends on Signal-to-Noise ratio of th data
- Decoding ๊ฒฐ๊ณผ ์กฐ๊ฑด A ๋ณด๋ค B์์ ํจํด์ด ๋ ๊ฐํ์ ๊ฒ์ด๋ค ๋ผ๋ ๊ฒ์, ๊ฒฐ๊ณผ๋ฅผ ๋ค๋ฅด๊ฒ ํํํ ๊ฒ ๋ฟ.
- ์ ๋์์ง๋ ์๊ธฐ ํ๋ค๋ค. Decoding์ ์ด๋ค ์ฐจ์ด๋ ๋ค ๋ฐ์ํ๊ธฐ ๋๋ฌธ์ด๋ค.
- Decoding accuracy๋ Signal-to-Noise์ ๋น์จ์ ์ํด ๊ฒฐ์ ์ด ๋๊ธฐ ๋๋ฌธ์,
- Signal is the distance between the mean values between conditions in the multidimensional space.
- Noise is the spread of the data within each condition.
- Lower decoding accuracy for one condition/group may be driven by the same level of single but higher level of noise.
Interpretation issue
- Through decoding analysis, we are looking at a correlation between brain activity and the stimulus category.
- Above-chance decoding means that 'any' differences in the stimulus category is correlated with the neural activity.
- ์ด๋ค ๋จ์ํ ์ฐจ์ด๋ decoding ๊ฒฐ๊ณผ์ ์ํฅ์ ์ค๋ค.
- ๊ทธ๋ฌ๋ ์ฃผ์๊ฐ ํ์ํ๋ค.
- When possible, decoding-irrelevant feature dimension of the stimulus should be controlled.
Understand complex human cognition
- ERP๋ก๋ถํฐ decodingํ ์ ์๋ ๊ฒ.
- What can we decode from ERPs.
- Location and Orientation
- The orientation from the previous trial
- Face identities and expressions
- motion directions
- positive vs negative valence of images
- letters
- standard ERP์ Decoding ๋ ์ฐจ์ด์ ์ด ์๋ค.
- How is decoding better/different from the standard ERPs analysis?
- Decoding is more sensitive than ERP analysis.
- Decoding is done separately for each subject, whereas the standard approach looks for consistency across subjects.
- Grater effect size.
- Limitations in the decoding analysis.
- 'any' differences between conditions potential produce above chance decoding.
- Decoding irrelevant feature dimensions should be controlled/counter-balanced.
'Biology > EEG' ์นดํ ๊ณ ๋ฆฌ์ ๋ค๋ฅธ ๊ธ
[ERP Decoding] EEG recording & Preprocessing (0) | 2023.02.12 |
---|---|
[ERP Decoding] ERP Decoding procedure (0) | 2023.02.11 |
[ERP Decoding] EEG - Electronencephalography (0) | 2023.02.07 |