Working Memory

Question 1

1. WM storage controversy

Answer 1

Early Views based on lesion studies: WM is primarily PFC

Problem of studies: activity during memory delay or memory impairment after lesions to a region do not necessarily mean that regions STORES the WM content  could just hold processes important to that storage

Both low-level and high-level cortices involved  distributed representation – but why?
• PFC could contain a duplication of sensory information
o Redundant, but might help in robustness against distraction
• However: differences in functional roles of regions
o Differing levels of abstraction in sensory regions (simple features) and PFC (abstract or verbal format)
o Differing functional roles of representation, from representation of sensory input to planning behavioural response
•  there seems to be a division of labour between low and high level regions
o Advantage: no necessity to duplicate low-level feature information in PFC
o PFC can then build flexible and task-dependent representations

Question 2

2. WM capacity controversy, slots vs resources

Answer 2

WM capacity is very limited
 There must be a restriction on the number of memories per node or the number of active nodes
 Is this restriction due to properties of individual storage regions or of the distributed network?

Competition for representation
• Multiple items compete within the same representational map, interfering with each other during encoding or delay period
• Evidence: visual WM capacity correlates with individual V1 volume, persistent stimulus selective activity per item decreases when number of items increases
• However: differing sensory modalities can impair each other, e.g. memorization of visual features reducing ability to retain auditory features

Distributed storage
• Items that are currently in the focus of attention are stored in a detailed way in the sensory cortex, secondary items for later use in a more abstract way in anterior regions
• Fits findings on neural activity

Task-relevant detail
• Level of detail required has implications for capacity
• Posterior to frontal gradient: complexity reduction, chunking
•  higher areas can retain larger chunks of low-level features

Question 3

3. Working memory active representation controversy (activity-silent WM)

Answer 3

What is the neurocomputational mechanism of retention? Two main theories:
• Activity-based retention
• Activity-silent retention
o = Stimulus-selective representations are maintained as a pattern of synaptic weights
• Might be a combination of both

Question 4

4. Persistent stimulus-selective activity

Answer 4

Necessary properties of regions encoding WM content:
• Stimulus selectivity: different memory contents should lead to different patterns of activity
• Persistent or delay-period activity: stimulus selective activity should persist over delay

Electrophysiological recordings on non-human primates
• Some findings on content specific delay period activity in the PFC
• But also in modality-specific sensory cortices
• Furthermore some other frontal, parietal and temporal areas

MVPA in humans
• Found delay-period stimulus specific activity in sensory cortical regions, frontal areas (e.g. FEF and PFC)
• Also EEG studies found activity in the PFC

 WM content is stored in a distributed way across sensory, parietal, temporal and frontal cortices

Persistent activity in the PFC
• reflects both retrospective memory and a prospective action plan
• Increases in strength until right before the response
• Should be highly task-specific

Question 5

5. Definitions of distributed representations

Answer 5

1. Local pattern information: referring to decoding stimuli from patterns of brain activity – distributed representation means information is distributed across a local population of neurons within one area
2. Separable information in multiple brain regions: multiple local response pattern exist in parallel and can be independently used to decode stimulus information
3. Inseparable information across multiple brain regions: information is encoded in global patterns of brain activity, and single areas alone would not allow decoding

How to disentangle this?
• Difference between 2 and 3 is the question whether information in multiple areas is redundant – does one region contain enough information already to decode a stimulus feature? Probably gradual rather than categorical
• Lesion studies
• Looking at choice probabilities: which signal is maximally predictive of a behavioural choice

Question 6

6. WM gradient

Answer 6

Look in the notes (graph).