Neuroscience of Working Memory
Serences et al. (2009, psychological science)
- Study to test the involvement of early visual areas in WM
- Task was for storage in WM of simple visual stimuli = contrast gratings
- Contrast gratings are used in many visual experiments because
representation of their characterisitics (e.g. orientation) relies on early visual
perception areas
- Serences et al. (2009) used contrast gratings as the memoranda for a
delayed response task, during fMRI scanning
- Participants were to remember the
orientation of the grating or the colour
- They were showed it for 1 second and
delayed it for 10 seconds
- They found that using a ‘typical’
univariate approach to fMRI analysis,
there was no sustained delay period
activity in primary visual cortex
sensitive to orientation or colour
- Nothing to do with working memory
Univariate analysis of fMRI data
- Recall that sustained delay period activity measured with fMRI was taken as
area for storage in WM
- Gold standard for a storage region = load-sensitivity (e.g., Leung & Postle) i.e.
a systematic increase in BOLD signal intensity with increasing memory load
- This is when fMRI BOLD signal is analysed with univariate statistics
- The typical approach to analysis has been to solve the general linear model
(GLM) in a univariate manner
1. Specify a model of expected responses to experimental manipulation (e.g.
load) – draw yourself what you think
the responses are going to look like
2. Fit the model to the BOLD data (lay
them over the top of each other)
3. Statistic is produced for goodness-
of-fit of model to data (regression) –
at a t-statistic of 6 is a very good fit
of the model
- brain is divided up into voxels (volumetric pixel) of ~mm3 each
- there are ~100k voxels in the brain – highly dimensional
- GLM is solved for each voxel independently
- Signal from neighbouring voxels are pooled and trials are averaged together
Limitations of univariate analyses
1
, - One voxel of 1mm3 contains ~700k neurons and they may not be doing the
same thing
- Pooling voxels in an area also assumes they are all doing the same thing
- Assumption that activity from pooled voxels are performing mental function
(e.g. short-term storage) that is independent of other parts of the brain
(modularity of brain function)
- Current theories on brain function propose that neural representations are
high-dimensional and supported by anatomically distributed computations
Information-based analyses
- This class of analysis looks at the patterns of activity of fMRI data – looks at
the brain as a whole and the voxel responses
- Instead of looking at statistical fit to a model, intensity of signal is ignored, and
the pattern produced by all the voxels is computed
- Information-based analyses are more sensitive to detecting information
representation in the brain, and are
more specific e.g. can distinguish
patterns between two visual objects in
the same brain area
- Lighter square = lower responses
- Darker square = larger responses
- You can now decode the information
that you are representing
Multi-voxel pattern analysis (MVPA)
- MVPA is a commonly used type of information-based analysis
- A classifier is first trained to identify patterns of voxels related to a specific
stimulus attribute or category e.g. a picture of a shoe vs a hat
- Once a classifier has learned this, it attempts to identify what the brain is
representing on an untrained data set
- For the Serences et al. (2009) experiment:
o Train classifier to distinguish between a contrast grating orientated at
45degrees vs 135degrees, or red vs green contrast grating
Scan someone doing a task and train a classifier to learn the
pattern of vowls in the brain
o Run the classifier on unseen
delayed response task data, in
which participants are retaining the
orientation of a contrast grating or
what colour is was:
o Classifier produces an accuracy score for that attribute i.e. how
accurately it was able to correctly identify the trained pattern of brain
activity – apply their learning to untrained data set
- In early visual areas not normally involved In memory, could decode the
information that was being retained in the delay period
2
Serences et al. (2009, psychological science)
- Study to test the involvement of early visual areas in WM
- Task was for storage in WM of simple visual stimuli = contrast gratings
- Contrast gratings are used in many visual experiments because
representation of their characterisitics (e.g. orientation) relies on early visual
perception areas
- Serences et al. (2009) used contrast gratings as the memoranda for a
delayed response task, during fMRI scanning
- Participants were to remember the
orientation of the grating or the colour
- They were showed it for 1 second and
delayed it for 10 seconds
- They found that using a ‘typical’
univariate approach to fMRI analysis,
there was no sustained delay period
activity in primary visual cortex
sensitive to orientation or colour
- Nothing to do with working memory
Univariate analysis of fMRI data
- Recall that sustained delay period activity measured with fMRI was taken as
area for storage in WM
- Gold standard for a storage region = load-sensitivity (e.g., Leung & Postle) i.e.
a systematic increase in BOLD signal intensity with increasing memory load
- This is when fMRI BOLD signal is analysed with univariate statistics
- The typical approach to analysis has been to solve the general linear model
(GLM) in a univariate manner
1. Specify a model of expected responses to experimental manipulation (e.g.
load) – draw yourself what you think
the responses are going to look like
2. Fit the model to the BOLD data (lay
them over the top of each other)
3. Statistic is produced for goodness-
of-fit of model to data (regression) –
at a t-statistic of 6 is a very good fit
of the model
- brain is divided up into voxels (volumetric pixel) of ~mm3 each
- there are ~100k voxels in the brain – highly dimensional
- GLM is solved for each voxel independently
- Signal from neighbouring voxels are pooled and trials are averaged together
Limitations of univariate analyses
1
, - One voxel of 1mm3 contains ~700k neurons and they may not be doing the
same thing
- Pooling voxels in an area also assumes they are all doing the same thing
- Assumption that activity from pooled voxels are performing mental function
(e.g. short-term storage) that is independent of other parts of the brain
(modularity of brain function)
- Current theories on brain function propose that neural representations are
high-dimensional and supported by anatomically distributed computations
Information-based analyses
- This class of analysis looks at the patterns of activity of fMRI data – looks at
the brain as a whole and the voxel responses
- Instead of looking at statistical fit to a model, intensity of signal is ignored, and
the pattern produced by all the voxels is computed
- Information-based analyses are more sensitive to detecting information
representation in the brain, and are
more specific e.g. can distinguish
patterns between two visual objects in
the same brain area
- Lighter square = lower responses
- Darker square = larger responses
- You can now decode the information
that you are representing
Multi-voxel pattern analysis (MVPA)
- MVPA is a commonly used type of information-based analysis
- A classifier is first trained to identify patterns of voxels related to a specific
stimulus attribute or category e.g. a picture of a shoe vs a hat
- Once a classifier has learned this, it attempts to identify what the brain is
representing on an untrained data set
- For the Serences et al. (2009) experiment:
o Train classifier to distinguish between a contrast grating orientated at
45degrees vs 135degrees, or red vs green contrast grating
Scan someone doing a task and train a classifier to learn the
pattern of vowls in the brain
o Run the classifier on unseen
delayed response task data, in
which participants are retaining the
orientation of a contrast grating or
what colour is was:
o Classifier produces an accuracy score for that attribute i.e. how
accurately it was able to correctly identify the trained pattern of brain
activity – apply their learning to untrained data set
- In early visual areas not normally involved In memory, could decode the
information that was being retained in the delay period
2
