Uncertainty and Bayesianism
Uncertainty. But why specifically Bayesian?
- Relevent for cognitive neuroscience
- Predictive processing
- A representation of information that we will build on in block 4.
Topic of today:
- How should we interpret probabilities?
- Where do the numbers come from?
- How should we update beliefs in light of evidence?
- Problems of logical omniscience and old evidence
- Issues for Bayesian confirmation: how to update knowledge in scientific sense?
- Normally: we do not reject a theory when it resists attacks (all or nothing)
- Bayesian: How likely is a theory true in light of evidence? But then: what role does old evidence play?
- Axiom: Maximum Entropy Principle
- Jeffrey's rule for updating
Starting exercise: what do we associate with Bayesianism? Different answers from different backgrounds?
- Bayes' rule
- Degrees of belief
- vs. frequentist
- predictive processing / Bayesian Brain
- Bayesian networks
- Probably someone from decision support systems
- Optimal models (mostly people from neuroscience angle )
- for how cognition ought to behave (normative rather than descriptive)
Interesting: dominant view is frequentist w.r.t. probabilities; they simply use Bayes' rule for updating.
Bayesianism according to Easwaran
- Prob. measures degree of belief
- not clear for freq. statistics: then prob. reflects frequency in the world.
- but this is the core of Bayesianism.
- For rational agents, degrees of belief must obey axioms of prob theory
- update beliefs by conditionalization
- Unique objective way to define a degree of belief
Objective Bayesianism: 1-4 Subjective: 1-3 Probabilism: 1-2
Axioms of probability theory
See other notes.
Defined over field of sets (powerset), the combination of every possible combination of subsets of a given amount of variables.
A,B,C binary variables, subsets are {A,B}, {A,C, etc. }
From these Kolmogorov axioms, all other probability theorems can be deduced.
E.g. what is probability of A and B, but their intersection is not empty (they do overlap).
- If prob of whole space is 1
- prob non negative
- so empty set has probability zero
We can derive from this:
- Pr(A or B) = P(A) + P(B) - (A and B)
- more general rule of axiom 3; in which case A union B is empty set.
Make drawing of intersection.
- P(B) = P(A union B) + P(B|A)
- P(A) = P(A union B) + P(A|B)
- they are disjoint, so we can add them together according to previous axiom "Pr(A and B)".
- P(A) + P(B) = P(A union B) + P(B|A) + P(A union B) + P(A|B) - P(A union B)
- =P(A union B) + P(B|A) + P(A|B)
Conditional probability?
W.r.t. conditional probability definition. Why does it semantically make sense?
Again: have a look at a square divided in two squares with overlap. E.g. for P(B|A): if you condition upon A, you "shrink" the universe to A, and then check in which section of that B also holds.
Mathematically convenient, but this cannot be derived from the Kolmogorov axioms! So it serves sort of as a fourth axiom!
Maximum Entropy Principle (Jaynes)
Solidified here.
Logic is objective but information is subjective.
So the probability of a state is always entirely conditional on a particular set of knowledge or information.
MEP: is the acknowledgment of our ignorance: I should maximize my ignorance about the things I have no information about. So limiting us to what we know, we should seek, given that constraint, the distribution with maximum entropy.
So MEP is a method for picking a prior distribution.
The hard part: what information do I, or some other observer, have? And then: do the maximization process.
Conditioning as update rule
Either using Bayes' rule, or its generalization (Jeffrey rule).
Bayes rule is easily mathematically derivable.
As an updating rule: prob. of new hypothesis is
People that accept this are Bayesians in a more broader sense.
Jeffrey: what if E is not observed with certainty? Can we still update?
Yes, with Jeffrey's rule. Takes into account the probability of the evidence.
P_new(H) = P(H|E=e) - P(E=e|H)xPr_old(H)\P(E=e) where, this is the crux, you do not known certaintly what P(E=e) is!
So you instead rewrite P(E=e) = sum over H, P(E=e | H=h1)P(H)
Reducible and irreducible certainty, example with tossing coins.
What if we do not know the likelihood of an event is? P(E|H)? E.g. in the case you do not know whether a coin is fair. In this case, you would have to do repeated experiments. How does your uncertainty about P(E|H) then change?
P(Heads)=0.5 in Bayesian account both when a) you know a coin is fair, and b) when you do not know if whether it is fair.
a) scenario with irreducible uncertainty; inherent to the stochastic nature of tossing a coin. * Risk * You will no longer learn in this scenario. b) scenario with reducible uncertainty; causes by ignorance about coins bias, but this uncertainty can be reduced by repeated observations! * Uncertainty * Allows for learning / updating belief.
Bayes' rule for updating is not useful in this scenario!
You do not have access the likelihood the event given your hypothesis P(E|H).
We can however use a binomial experiment over H, with a prior parameter q. Let q stand for probability of heads.
So what instead of treating q as a number, e.g. 0.5 Instead, we can define a distribution of q, which is called a hyperprior. The parameters of this hyperprior is a hyperparameter.
In case of coin tosses, we can estimate the success of n flips using the binomial distribution (in this case we have an exact formula).
For binomial experiment, we have for P(H)=q the hyperprior beta distribution, with hyperparamters alpha and beta.
Sketch cases for a=1,b=1: with 0 successes and failures, distribution is uniform.
Collaps hyperparameter into expected P(E|H): P(heads) = a / (a b), in that case 1/2 (without further observation).
a =2, b=1, linear line; a (heads) more likely. P(heads)=2/3.
a =2, b=2, exponential.
Still P (heads)=1 / 2, but with more certainty.
Paradoxes and problems
- Logical omniscience
- Problem of old evidence
- Bertrands' Paradox
Logical Omniscience
Bayesian Confirmation Theory (BCT) vs hypothetico-deductive model of confirmation; in the latter case you do not assign credence, you only show that the theory is not (yet) rejected.
BCT also gives a indication of how much better one hypothesis is over another.
An agent "knows" everything that it can logically infer from its axioms.
Likewise, the agent knows about P(X) for all relevant variables. Completely unrealistic for a model of any cognitive agent.
(But this problem is a problem of epistemic logic in general; not specifically BCT )
Old Evidence
Old evidence can confirm hypothesis.
E.g. the perihelion shift of Mercury (in an old theory) also evidence for general relativity.
If we assume the evidence is already known, then P(E=e)=1, so P(H|E)=P(H), i.e. the posterior is equal to the prior! The evidence does not, probabilistically speaking, support for the new theory. But intuitively, it is. So BCT seems to be lacking here.
Does BCT not fit in this scenario?
Solutions? P(E=e) not 1, but close to 1? Would still have a minimal impact.
Bertrand's Paradox
Cubical box has side length between 1 and 3 meters.
How to assign subjective beliefs about the area of the box?
Let's say P(l in [1,2]) = 0.5 and P(l in [2,3])=0.5 Which is perfectly reasonable.
See page 6/12, or 326.
Paradox: if you assign belief on sides, belief on area is non-uniform, and vice versa when you assign belief on area.
The point: there is no universal right way to define a prior.
Many papers on this problem, could be topic for specialization project.