Neural Networks and Cognitive Control (2) Flashcards
What is cognitive control?
- effects perception, action, attention, long-term memory, language and decision making, but is not exactly each of these things
- guides and coordinates all of these processes in services of our plans and goals
What is cognitive control also referred to as?
- executive control
- central executive
- self control
What is an early hypothesis about what is in control? What is the issue with this?
- homunculus: mini me
- but what is going on in mini-me’s head and who is making it’s decisions
- never-ending homunculi
What is meant by “banish the homunculus”?
- stop appealing to an ill-defined and circular self, central executive or consciousness to explain goal-oriented behaviour
What is the goal of figuring out what is in control?
- a mechanistic account of cognitive control in neural, psychological and computational terms
What are the components of cognitive control?
- shifting between tasks or mental sets
- updating and maintaining working memory representations
- inhibition of dominant or prepotent responses
- monitoring and adjusting performance
- etc.
What is an example of shifting between tasks?
- local-global task
- presented with letters made up of smaller letters
- if red, report global letter
- if blue, report local letters
- must shift between tasks depending on colour
What is an example of updating and maintaining working memory representations?
- letter memory task
- given a list of words and once the list stops need to report the last 4 words
- need to continual update memory to do task
What is an example of inhibition of dominant responses?
- stroop task
- presented with colour words in different colours of ink
- must either report color word or ink color
- must inhibit dominate task of reading word in order to report ink colour
What brain area(s) are involved in cognitive control?
- in general Prefrontal cortex
- Lateral: “dorsolateral”! (DLPFC), ventrolateral
- Ventromedial: orbital frontal
- Medial: dorsomedial, “anterior cingulate”! (ACC)
What is the stroop task and guided activation theory model?
- feed forward network
- input: stimulus features of colour and word
- hidden layers
- output: responses: “red” or “green”
- hidden layers also input to control representations of colour or word
What are the conditions of the stroop task?
- control: colour naming without word and word reading without colour
- conflict: color naming and word reading with conflicting colours
- congruent: colour naming and word reading with congruent colour
What does the empirical data of the stroop task demonstrate?
- word reading was fast for all conditions
- colour naming was significantly slower for conflict conditions and slightly faster for congruent conditions
What are the layers of the neural network model for the stroop task?
- visual perception
- perceptual-motor mapping
- verbal response
- goal maintenance/cognitive control
What are the stronger pathways in the neural network model for the stroop task?
- shown with darker black lines
- reporting words are strongest from input to hidden layer and hidden layer to verbal response
- reading green is easy/fast because it is one, highly practiced response
How does the control condition correspond to the neural network model of the stroop task?
- control word: easy/fast, only one highly practiced response
- control ink: pretty easy/medium fast, one somewhat practiced response
How does the incongruent condition correspond to the neural network model of the stroop task?
- incongruent word: easy/fast, two competing response but highly practiced response wins
- incongruent ink: hard/slow, two competing responses
- requires top-down cognitive control to inhibit word reading and facilitate colour naming
What does simulation data of the stroop task model show?
- confirms the trends of the empirical data
How is top-down control implemented?
- goal representations in dorsolateral prefrontal cortex (DLPFC) provide top-down biasing of perceptual-motor mappings
- working memory provides the representation
- inhibition is the resulting process
What is cognitive control overall?
- cognitive control is not a set of independent faculties, but rather an integrated system
What is the Eriksen flanker task?
- central letter is surrounded by “flankers”
- flankers can be congruent or incongruent (HHHHH, HHSHH)
What were the findings of the Eriksen flanker task?
- flanker effect: incongruent trials are slower than congruent trials
- Gratton effect: the flanker effect is smaller after incongruent trials then after congruent trials
Why is the gratton effect seen?
- after an incongruent task the level of cognitive control is turned up which makes a subsequent incongruent task faster
Why is the flanker effect seen?
- incongruent trials require top-down control
What additional part of the neural network was identified with the eriksen flanker task model?
- ACC involved in conflict monitoring
- when conflict detected, it sends signal to DLPFC to amp up control
- either enhances or inhibits L and R or C
What does the model of the eriksen flanker task model show?
- demonstrates the same data
- flanker effect and gratton effect
How is self-regulation of control implemented?
- anterior cingulate cortex (ACC) monitors performance for response conflict and signals dorsolateral prefrontal cortex to provide increased top-down biasing of perceptual-motor mappings
What is the 1-2-AX task?
- presented with a sequence of letters and digits
- press left button for each except
- press right button for X preceded by A, if most recent number was 1
- press right button for Y preceded by B, if most recent number was 2
What is required of someone to do the 1-2-AX task?
- maintain multiple items
- update specific items
- ignore other items
- use memory to control performance
- learn when/what to maintain, store, ignore, forget
What circuitry is used for the 1-2-AX task?
- circuitry of cortico-basal ganglia-thalamocortical loops
- prefrontal cortex, basal ganglia working memory (PBWM) model
What gating is important for the 1-2-AX task?
- dopamine gating of working memory
- the gate is open when working memory is being updated
- the gate is closed when working memory is maintaining
How does reinforcement learning take place in the 1-2-AX task model?
- reinforcement learning from dopamine prediction-error signal
How do we learn how to control?
- working memory updating and maintenance controlled by cortico-basal ganglia-thalamocortical loops
- phasic dopamine signals indicate reward prediction errors
- dopamine signal used to learn how to maximize reward by gating relevant information into working memory
- controlling working memory maintenance and updating = internally directed action control
What are the models of cognitive control?
- hierarchical
- Botvinick: control actions based on low-level or high-level interpretations (ex. which finger to respond, leave lab or stay)
- can be physical or mental (ex. make coffee -> need sugar and milk -> add sugar..hold..pick up)
- Fuster’s hierarchy: motor starts with high level goals and breaks it down to lower level goals, higher levels apply top-down control
Is there hierarchical organization of the prefrontal cortex?-
- higher levels moving towards the front of the brain
- still not well understood
- Frontopolar cortex: tracking goals
- mid-dorsolateral PFC: task switching
- posterior lateral PFC: selection of sequences of responses
- Premotor cortex: response selection
What are the key components of control?
- top-down
- self-regulated
- learns
- hierarchical
