IBSC Seminar on Priming

1
IBSC Seminar on Priming
Steve Gotts CNBC and NIMH/NIH
2
Overview
I) Review of empirical data on priming and
related neural changes II) Discuss issues
raised in last IBSC meeting A) Are
different forms of priming associated with the
same neural effects? B) What might be
done to evaluate the nature of
representations used in connectionist
models? III) Implications for connectionist
models
3
I) Review of empirical data on priming and
related neural changes

• Behavioral priming - change in the speed, bias,
  or accuracy of the processing of a stimulus,
  following prior experience with the same or a
  related stimulus
• indirect vs direct tasks
• variety of indirect tasks (stem completion,
  naming, LD, etc.)
• component process view
• multiple processes contribute to any given task
• response time is an aggregate measure of
  facilitation
• occurring for each component processes
• priming is greatest when processes engaged at
  prime and
• probe match

4

• Repetition Suppression - decrease in hemodynamic
  or neural activity following repetition of the
  same or a related stimulus
• Empirical generalizations
• regions showing suppression are normally
  restricted to those
• that are responsive to the type of stimuli
  being used
• suppression is observed in multiple brain
  regions, suggesting
• that changes can be observed at multiple
  functional loci
• (like the component process view of priming)
• suppression is not observed in all regions
  associated with
• processing stimuli in a particular task (e.g.
  not often observed
• in primary sensory or motor regions, at least
  in fMRI and PET)

5
(No Transcript)
6
Priming as a memory phenomenon

• dissociations between implicit and explicit
  tasks (e.g. amnesia)
• explicit retrieval is often associated with
  enhanced rather than
• decreased neural activity in medial temporal
  and prefrontal
• regions
• studied in a variety of tasks/paradigms
• - word-stem completion (implicit -gt decreases)
• - conceptual tasks (generally decreases in left
  inferior
• frontal and ventral occipitotemporal)
• - comparisons of implicit and explicit tasks
  (explicit task
• on probe can reduce or reverse the
  decreases)
• - masked priming (sometimes decreases, sometimes
  
• increases increases in gradual unmasking and
  with
• backward mask of probe)

7
Priming as a tool for studying representations

• fMR Adaptation (Dehaene, Grill-Spector) vary a
  sequence of
• stimuli along a single stimulus dimension in
  order to measure the
• sensitivity of particular cortical regions to
  that dimension (e.g.
• vary object viewpoint, size, position, etc.)
• Hyper-resolution using the adaptation technique
  may afford
• within-voxel discrimination by mean-activity
  level

8
(No Transcript)
9
Grill-Spector Malach (2001)
10
Priming as a tool for studying representations

• Familiar vs unfamiliar object priming
• Repetition increases are often observed for
  unfamiliar stimuli, whereas decreases are
  observed for familiar stimuli
• (e.g. Henson et al., 2000 Schacter et al., 1995)
• Henson's proposal
• regions that show repetition enhancement are
  those that subserve
• a process that occurs only on the probe and not
  the prime
• regions that show repetition suppression
  subserve processes
• operating on both prime and probe

11
(No Transcript)
12
Priming as a tool for studying representations

• Lag effects
• suppression (and enhancement) attenuate with
  lag from 10 sec to
• 20 min (2 to 140 intervening stimuli)
  (Henson et al., 2000)
• both lag and intervening stimuli attenuate
  suppression in
• occipitotemporal, with a progression to
  longer-lived effects as one
• moves from posterior to anterior (Henson et
  al., in prep)
• priming/suppression in object naming is greater
  at 30 sec
• compared to 3 days, but is significant at both
  (van Turennout et
• al., 2000)

13
Priming as a model domain for relating mind and
brain

• sharpening theory of Desimone (1996) Wiggs
  Martin (1998)
• short-term adaptation in fMR adaptation
• some relevant single-cell physiology studies in
  monkeys
• McMahon Olson (SFN 2003, 2004)
• Baker, Behrmann, Olson (2002)
• Rainer Miller (2000) Freedman et al. (SFN
  2004)
• Li, Miller, Desimone (1993)
• Miller, Li, Desimone (1993)
• long-term effects of practice are consistent
  with "sharpening"
• short-term effects of repetition are consistent
  with local adaptation
• or other negative feedback mechanisms

14
(No Transcript)
15
II) Discuss issues raised in last IBSC meeting

• A) Are different forms of priming associated with
  the same
• neural effects?
• 1) Categorical (duck-chicken) vs. Associative
  (coat-rack)
• Kotz et al. (2002) fMRI
• auditory lexical decision with pairs of
  categorically or
• associatively related words
• decreases in left inferior frontal gyrus
• increases in posterior middle temporal cortex
• greater activity to categorical pairs than to
  associative pairs in
• posterior medial parietal/cingulate

16
II) Discuss issues raised in last IBSC meeting

• A) Are different forms of priming associated with
  the same
• neural effects?
• 2) Strategic priming effects
• Mummery et al. (2002) PET
• study of semantic priming using lexical
  decision and varying
• relatedness proportion
• prime word presented for 50 ms, followed by
  target word
• trend for greater semantic priming with higher
  relatedness
• proportion
• correlated with greater decreases in left
  anterior temporal and
• anterior cingulate (although 100 gt 75 in
  temporal)

17
II) Discuss issues raised in last IBSC meeting

• A) Are different forms of priming associated with
  the same
• neural effects?
• 3) Expectation effects
• Jiang et al. (2000) fMRI
• DMS task with faces targets and distractors
  could repeat
• the first target post-sample elicited enhanced
  activity in
• ventral temporal and frontal/insular cortex
• subsequent repetitions of the target decreased
  in ventral
• temporal, but not in frontal/insula
• repeated distractors elicited suppressed
  activity in ventral
• temporal regions

18
II) Discuss issues raised in last IBSC meeting

• B) What might be done to evaluate the nature of
• representations used in connectionist
  models?
• fMR adaptation could probably be used
  productively to evaluate the correspondence
  between distributed representations in
  connectionist models and real neurons
• recently used to plot detailed tuning curves for
  number in parietal cortex can distinguish
  between log and Guassian shapes S.Dehaene
• orthography
• phonology
• semantics
• caveat need to control for strategic processing
  to the
• extent possible (masking and short delays?)

19
III) Implications for connectionist models
Bottom line Most connectionist models will NOT
show repetition-related decreases in unit
activity So What is currently missing
from connectionist models, and how badly
does it matter?
20
What is currently missing from connectionist
models?
.. or alternatively, why do the changes happen in
real neurons?

• 1) for short-term effects (lt 1 s) of repetition
  suppression/priming,
• firing-rate adaptation?
• synaptic depression?
• priming could be due to residual activity
• but may create difficulties for learning
  representations
• 2) for slightly longer effects (1 s - 1 min),
  adaptation and synaptic
• depression can still work for the activity
  changes, but
• accounting for the priming effects gets
  harder
• residual activity is less tenable
• firing-rate decreases can be greatest for the
  "best" cells
• changes to long-term synaptic strengths might
  help, though

21
Stimuli that lead to larger firing rates tend to
produce larger repetition suppression effects
(Miller et al., 1993 Li et al., 1993)
Adaptation produces this effect in a
connectionist model
2/3 of cells that were visually responsive
showed match-nonmatch differences
1/4 of cells that were visually responsive
showed match-nonmatch differences
Compression of firing rates is the opposite of
representational sharpening
vs.
22
What is currently missing from connectionist
models?

• 3) For long-term effects (minutes, hours, days,
  ..), need to
• figure out what changes to learning rules can
  produce a
• progressive "sharpening" of activity, while
  decreasing
• overall activity
• this would be consistent with several
  single-cell
• recording studies (Rainer Miller, 2000
• Baker, Behrmann, Olson, 2002 Freedman et
  al., 2004)

All learning rules that actually learn the
patterns increase response selectivity Not all
of them lead to average decreases, leaving the
peak activity unchanged