Information Integration and Decision Making in Humans and

Download Report

Transcript Information Integration and Decision Making in Humans and

Perceptual Inference and Information
Integration in Brain and Behavior
PDP Class
Jan 11, 2010
How Neurons in Perceptual Systems Might
Carry Out Perceptual ‘Inferences’
• Each neuron (or collection of neurons) is treated as
standing for an hypothesis about what is out there in
the world:
– An oriented line segment at a particular point in space
– Something moving in a certain direction
– A monkey’s paw
• Note that a given object or scene might be
characterized by a number of hypotheses; there might
or might not be a separate ‘grandmother’ hypothesis.
• We treat the firing rate of each neuron as corresponding
to the degree of belief in the hypothesis it participates
in representing, given the available evidence, expressed
mathematically as P(H|E)
• Question: How can we compute P(H|E)?
Example
• H = “it has just been raining”
• E = “the ground is wet”
• Assume we already believe:
– P(H) = .2; P(~H) = .8
– P(E|H) = .9; P(E|~H) = .01
• Then what is p(H|E), the probability that it has
just been raining, given that the ground is
wet?
Theory of Perceptual Inference:
How Can we Compute p(H|E)?
• Bayes’ Rule provides a formula:
P(H|E) =
where
p(E|H)p(H)
p(E|H)p(H) + p(E|~H)p(~H)
– P(H) is the prior probability of the hypothesis, H
– P(E|H) is the probability of the evidence, given H
– P(~H) is the prior probability that the hypothesis is false
(and is equal to (1-P(H))
– P(E|~H) is the probability of the evidence, given that the
hypothesis is false.
• Bayes rule follows from the definition of
conditional probability
Derivation of Bayes Rule
•
•
•
•
p(H|E) = p(H&E)/p(E)
p(E|H) = p(H&E)/p(H)
So p(E|H)p(H) = p(H&E)
Substituting in the first line, we obtain
p(H|E) = p(E|H)p(H)/p(E)
• What is p(E)?
P(E) = p(H&E) + p(~H&E)
= p(E|H)p(H) + p(E|~H)p(~H)
Example
• H = “it has just been raining”
• E = “the ground is wet”
• Assume we believe:
– P(H) = .2; P(~H) = .8
– P(E|H) = .9; P(E|~H) = .01
• Then what is p(H|E), the probability that it has
just been raining, given that the ground is
wet?
(.9*.2)/((.9*.2) + (.01*.8)) = (.18)/(.18+.008) = ~.96
• What happens if we change our beliefs about:
– P(H)? P(E|H)? p(E|~H)?
How Should we Combine Two or
More Sources of Evidence?
• Two different sources of evidence E1 and E2 are
conditionally independent given the state of H, iff
p(E1&E2|H) = p(E1|H)p(E2|H)
p(E1&E2|~H) = p(E1|~H)p(E2|~H)
• Suppose p(H), p(E1|H) and p(E1|~H) are as before and
E2 = ‘The sky is blue’; p(E2|H) = .02; p(E2|~H) = .5
• Assuming conditional independence we can substitute
into Bayes’ rule to determine that:
p(H|E1&E2) =
.9 x .02 x .2
= .47
.9 x .02 x .2 + .01 x .5 X .8
• In case of N sources of evidence, all conditionally
independent under H, then we get:
p(E|H) =
Pj p(Ej|H)
Conditional and Unconditional
Independence
•
Two variables (here, x and y) are ‘(unconditionally) independent’ iff
p(x&y) = p(x)p(y) for all x,y.
•
Two variables are ‘conditionally independent’ given a third (z) iff
p(x&y|z) = p(x|z)p(y|z).
•
The variables x and y are unconditionally independent in one of the
graphs above. In the other graph, they are conditionally independent
given the ‘category’ they are chosen from, where this is represented
by the symbol used on the data point, but they are not
unconditionally independent.
How this relates to neurons
•
•
•
•
•
•
It is common to consider a neuron to
have an activation value corresponding
to its instantaneous firing rate or
p(spike) per unit time.
The baseline firing rate of the neuron is
thought to depend on a constant
background input called its ‘bias’.
When other neurons are active, their
influences are combined with the bias to
yield a quantity called the ‘net input’.
The influence of a neuron j on another
neuron i depends on the activation of j
and the weight or strength of the
connection to i from j.
Note that connection weights can be
positive (excitatory) or negative
(inhibitory).
These influences are summed to
determine the net input to neuron i:
neti = biasi + Sjajwij
where aj is the activation of neuron j,
and wij is the strength of the connection
to unit i from unit j.
Input from
neuron j
wij
Neuron i
A Neuron’s Activation can Reflect
P(H|E)
•
•
The activation of neuron i given its
net input neti is assumed to be
given by:
ai = exp(neti)
1 + exp(neti)
This function is called the ‘logistic
function’.
•
It is easy to show that
ai = p(Hi|E) iff
aj = 1 when Ej is present,
or 0 when Ej is absent;
wij = log(p(Ej|Hi)/p(Ej|~Hi);
biasi = log(p(Hi)/p(~Hi))
•
In short, idealized neurons using
the logistic activation function can
compute the probability of the
hypothesis they stand for, given
the evidence represented in their
inputs, if their weights and biases
have the appropriate values.
ai
neti
Accurately Coding Probability in a
Short Interval of Time
• If p(spike per 10 msec) = p(H|E) then having a single
neuron to represent a hypothesis would make it difficult
to get a clear estimate of P(H|E) within, say, 100 msec.
• However, suppose many (say, 10,000) neurons each
encode the same hypothesis, and suppose that they
produce spikes independently of each other (but based
on the same p(H|E)).
• Then the number of spikes summed over the population
would provide a very close approximation of p(H|E)
even in a brief interval such as 10 msec.
Information Integration in Human Perception:
The McGurk Effect (McGurk & MacDonald, 1976,
Nature 264, 746-748)
• First listen to clip with
your eyes closed.
• Then listen again with
eyes open.
• What you see appears to
influence what you hear.
• What your hear probably
sounds like ‘ba’ by itself.
• What does it sound like
when you watch the
face?
• Most people hear ‘da’ or
‘tha’.
• McGurk effect movie from
USC.
•
•
•
•
•
Massaro et al (2001) performed
an experiment in which subjects
received auditory inputs ranging
from a good “ba” sound to a good
“da” sound and visual speech
inputs ranging from a good “ba”
to a good “da”.
The results are consistent with
the model we have been
describing, with auditory and
visual input treated as
conditionally independent sources
of evidence for the identity of the
spoken syllable.
Note that when the auditory input
is at either extreme, the visual
input has little or no effect.
These are examples of ‘floor’ and
‘ceiling’ effects that are often
found in experiments.
The model explains why the effect
of each variable is only found at
moderate values of the other
variable.
Application of the
model to a McGurk
experiment
Choosing between N alternatives
•
•
•
•
Often we are interested in cases where
there are several alternative
hypotheses (e.g., different directions of
motion of a field of dots). Here we
have a situation in which the
alternatives to a given H, say H1, are
the other hypotheses, H2, H3, etc.
In this case, the probability of a
particular hypothesis given the
evidence becomes:
P(Hi|E) =
p(E|Hi)p(Hi)
Si’p(E|Hi’)p(Hi’)
The normalization implied here can be
performed by computing net inputs as
before but now setting each unit’s
activation according to:
ai = exp(neti)
Si’exp(neti’)
This normalization effect is
approximated by lateral inhibition
mediated by inhibitory interneurons
(shaded unit in illustration).
H
E
‘Cue’ Integration
in Monkeys
• Saltzman and Newsome (1994)
combined two cues to the
perception of motion:
– Partially coherent motion in a specific
direction
– Direct electrical stimulation of neurons
in area MT
• They measured the probability of
choosing each direction with and
without stimulation at different
levels of coherence (next slide).
Model used by S&N:
• S&N applied a model that is
structurally identical to the one we
have been discussing:
– Pj = exp(yj)/Si’exp(yj’)
– yj = bj + mjzj + gdx
– bj = bias for direction j

mj = effect of micro-stimulation
– zi = 1 if stimulation was applied, 0
otherwise

gd = support for j when motion is in that
direction (d=1) or other more disparate
directions (d=2,3,4,5)
– x = motion coherence
• Open circles above show
effect of presenting visual
stimulation in one direction
(using an intermediate
coherence) together with
electrical stimulation
favoring a direction 135°
away from the visual
stimulus.
• Dip between peaks rules
out simple averaging of the
directions cued by visual
and electrical stimulation
but is approximately
consistent with the
Bayesian model (filled
circles).