Introduction to Computational Neuroscience Lecture 11: Attention & Decision making
Lesson Title 1 Introduction 2 Structure and Function of the NS 3 Windows to the Brain 4 Data analysis 5 Data analysis II 6 Single neuron models 7 Network models 8 Artificial neural networks 9 Learning and memory 10 Perception 11 Attention & decision making 12 Brain-Computer interface 13 Neuroscience and society 14 Future and outlook: AI 15 Projects presentations 16 Projects presentations Basics Analyses Models Cognitive Applications
https://www.youtube.com/watch?v=vjg698u2mvo https://www.youtube.com/watch? v=igqmdok_zfy&feature=player_embedded http://channel.nationalgeographic.com/channel/brain-games/videos/ brain-games-pay-attention/
Learning objectives Recognize the role of attention in shaping perception, memory, and decision making Understand basic theoretical framework and neuronal models of decision making
Definition Attention: process by which some information is selected for further processing and other information is discarded Assumption: brain does not have the capacity to fully process all the information it receives. Attention is a filter or bottleneck in processing
Types Several types: e.g. spatial attention, feature-based attention Spatial attention implements a sort of spotlight, highlighting locations, moving (searching), zooming in or out
Visual spatial attention In vision, the locus of the spotlight can be the same as eye fixation (overt attention) but it doesn t have to (covert attention) Limit of the metaphor: attention can be split in 2 locations http://ocw.mit.edu/courses/brain-and-cognitive-sciences/9-00scintroduction-to-psychology-fall-2011/attention/discussion-attention/
Visual spatial attention Exogeneous orienting: external cue
Visual spatial attention Endogeneous orienting: task demand
Saliency maps How does the spotlight know where to go? Models postulate the existence of a saliency map, which would represent topographically the relevance of different parts of the visual field Could be built only based on bottom-up information (e.g. how different is a stimulus compared to its neighborhood?) or could include task dependent (top-down) information There is some evidence that neurons in LIP could encode saliency
Saliency maps
Posner s task * focus visual attention to an area by using a cue, e.g. arrow * measure reaction time to detect target when: i) observer doesn t know where item will appear (neutral cue) ii) observer is cued to where item will appear (valid cue) iii) observer is wrongly cued (invalid cue)
Posner s task Detection is faster for valid targets Detection is slower for invalid cues (inhibition of return)
Where is Wally?
Search tasks Task of detecting the presence or absence of a specified object (target) in the middle of other objects (distractors) * feature search: find the red cross, find the circle * conjuction search: find the orange square
Feature Integration Theory In feature search: RTs do not increase as a function of the display size as they do in conjunction search In conjunction search: the binding problem has to be solved Interpretation: 2 processes * a first pre-attentive stage, processes that are fast, parallel and involuntary (pop-out) * a second stage, attentive, serial. Volitional deployment of attention
Where is Wally?
Blind out of focus of attention? Change blindness: unless we attend an object, we are unlikely to perceive consciously in any detail and detect when it is changed http://ocw.mit.edu/courses/brain-and-cognitive-sciences/9-00sc- introduction-to-psychology-fall-2011/attention/discussion- attention/ But we are not totally blind outside the focus of attention (negative priming)
Blind out of focus of attention? Attentional blink: we tend not to notice that happens shortly after something else happens (temporal limitation)
Attention and appearance Spatial attention improves detection rate, and reaction times, and also discrimination performances Attention enhances apparent contrast, perceived motion coherence, and spatial frequency Attention changes perceived size and makes moving objects appear to move faster Thus attention not only enhances perception, it also distorts our representation of the visual scene according to the behavioral relevance of its components
Neural basis Where does attention come from? Which parts of the brain does it affect? What does it change in the neuronal response? - amplitude? - tuning? - baseline? - noise? - temporal properties? Can we explain the perceptual changes based on the neuronal changes?
Changes depends on... * cortical area * neurons * task demands
Biased competition Multiple stimuli in the visual field activate population of neurons that engage in competition Ex. area V4, 2 stimuli in neuron s RF After some delay, the pair response is almost entirely driven by the attended stimulus
Gain modulation
Baseline changes Increase in spontaneous activity before the stimulus was presented at the cued location fmri, increases in all visual areas but stronger in V4
Changes in oscillations
Decision making To survive an thrive, all animals must derive knowledge about the world from sensory observations shuttlecock landing? predator in the high grass?
Decision making An observer s knowledge of a world state can be expressed as a probability distribution (predator presence, landing location,...) Knowledge is not sufficient for organisms; actions are needed!
Decision making Volvariella volvacea Amanita phalloides
Decision making Would you eat it?
Decision making (theory) Decision making can be thought of as a form of statistical inference decide = select among competing hypotheses h1, h2 (or more) Elements of this decision process: - Evidence: - Prior information: - Cost (or value) function: e.g. Amanita rarely grows in the area eating Amanita = -infty eating Volvariella = +10 not eating = -5
Decision making (theory) 1) Bayes theorem can be used to combine evidence and prior information to infer the probability of different states of the world (s) based of observation (x) 2) A posteriori expected loss to be optimized is
Classification (0/1 loss func.) For classification tasks we can consider the loss function: for which Bayes optimal decision rule would be
Classification (0/1 loss func.) Using Bayes: clarifies how evidence and prior information is combined
Classification (0/1 loss func.) To decide we compare posterior probabilities. Choose s1 if: P (s 1 x) = f(x s 1)P (s 1 ) f(x) > P (s 2 x) = f(x s 2)P (s 2 ) f(x)
Classification (0/1 loss func.) Just reorganizing the terms of the inequality: f(x s 1 ) f(x s 2 ) > P (s 2) P (s 1 ) This is known as the likelihood ratio test = optimal decision rule If the prior probabilities are equal (0.5), choose s1 if: LR = f(x s 1) f(x s 2 ) > 1
Classification (general) In general, we consider a loss function matrix L(si,sj) that specifies the cost associated with choosing sj when the true class was si S = { disease, no disease } A = { surgery, no surgery } L( disease, surgery ) = 0 L( no disease, no surgery ) = 0 L( disease, no surgery ) >> L( no disease, surgery ) LR = f(x s 1) f(x s 2 ) >t
Decision variables LR (or any monotonic function of it) provides an optimal decision variable This framework can be extended to the situation where we have multiple pieces of evidence observed over time. LR accumulates the evidence in time LR = nx i=1 log f(x i s 1 ) f(x i s 2 ) When the decision variable reaches a threshold t (possibly reflecting prior information and values) a decision is made
Random walk model Related to this framework are the random walk and race models of decision making developed by psychologists to explain behavioral data The decision variable is the cumulated sum of the evidence. The bounds represent the stopping rule
Random walk model Well studied mathematically (diffusion process) Many variants (discrete time, continuous time, leaky integrate) Successfully account for: - Distribution of reaction times - Speed-accuracy tradeoff (decreasing the threshold has the effec of increasing speed and decreasing accuracy)
Race model Another variant is the race model Two or more decision processes represent the accumulated evidence for each alternative
Anything like that in the brain?
Yes Study decisions on perceptual tasks Mike Shadlen, Paul Glimcher (and others)
?
https://www.youtube.com/watch?v=odxcytn-0os
Direction motion task Monkey decides between 2 possible opposite directions, and saccade to signal his choice whenever he is ready Task difficulty is controlled by varying the coherence level Decision = problem of movement selection
Accumulation of evidence (LIP) LIP receives inputs from MT and MST, outputs in FEF and SC (generation of saccades) LIP is implicated in selection of saccade targets, working memory, intention, etc... Record neurons which have one of the choice targets in the response field and the other outside
Accumulation of evidence (LIP) After 220 ms, response reflects decision - faster rise for easier choices, decrease for opposite direction Aligning responses to saccade initiation reveals correlate of commitment: a threshold rate at which decision is made, 70 ms before saccade initiation
Accumulation of evidence (LIP)
Accumulation of evidence (LIP) Responses achieve a common level of activity, 70 ms before saccade When monkey chooses direction, another set of neurons behave similarly As if the fact that they reach a threshold value determines the termination of the decision process
Accumulation of evidence (LIP) Pattern of LIP activity matches prediction of diffusion/race models Rise of activity appears to reflect accumulation of evidence Evidence could come from a difference in activity of pools of MT neurons with opposite direction preferences, which was suggested to approximate log(lr) LIP activity = decision variable? neurons implement log(lr) test? How is the threshold set?
Accumulation of evidence (LIP) Changing the reward associated with each target (value) Vary the probability that a saccade to a target will be required (prior) Offset (distance to decision threshold) of the responses of LIP neurons are modified by reward and priors Studies suggest that behavioral value of a response and priors are encoded in LIP (offset)
Neuroeconomics
Summary Attention refers to the process by which some information is selected for further processing, shaping perception and memory Decision making is a process that weights (Bayesian?) priors, evidence, and value to generate an action Accumulation of evidence and comparison to a given threshold seems to be a basic mechanism of neuronal decision making
Lesson Title 1 Introduction 2 Structure and Function of the NS 3 Windows to the Brain 4 Data analysis 5 Data analysis II 6 Single neuron models 7 Network models 8 Artificial neural networks 9 Learning and memory 10 Perception 11 Attention & decision making 12 Brain-Computer interface 13 Neuroscience and society 14 Future and outlook: AI 15 Projects presentations 16 Projects presentations Basics Analyses Models Cognitive Applications