Inferring something unknown
Difficult to infer unknown target
f
outside
How to infer the orange probability?
Sampling
- Bin: ASSUME orange prob =μ , green prob =1−μ with μ unknown
- Sample:
N
marbles sampled independently, with orange fraction
=ν , green fraction =1−ν , with ν known.
Possible vs. Probable
- possibly not: sample can be an extreme case, so we cannot possibly learn f
- probably yes: in-sample
ν likely close to unknown μ
Hoeffding’s Inequality
With a sample size of
N
,
ℙ[|ν−μ|>ϵ]≤2exp(−2ϵ2N)
The probability of big deviation( ϵ ) is small.
The statement ν=μ is probably approximate correct(PAC).
- valid for all
N
and
ϵ - doesn’t depend on μ , so we do not need to know μ
- larger size of
N
or loose gap
ϵ⇒ higher probability for μ≈ν
If larger N , can probably infer unknown
μ from ν