2
$\begingroup$

Bits Back coding is a scheme to transmit an observation $x$. You can read about it here[1]. To my understanding, it works like this:

  1. The encoder samples a message $z$ from a distribution $Q(z|x)$ that it computed.
  2. The encoder sends $z$ to decoder, using a code based on $P(z)$, known to both.
  3. The decoder recovers $z$ and computes $P(x|z)$.
  4. The encoder sends $x$ to the decoder using a code based on $P(x|z)$.
  5. The decoder recovers $x$.

In reality, the messages in steps 2 and 4 can be combined. Naively, the cost of this protocol would be the expected number of bits in these messages.

However, the trick is that once the decoder recovered $x$, it can compute $Q(z|x)$ using the same algorithm the encoder used, and then recover all the auxiliary bits it had to use internally in order to sample $z$ from $Q$. The expected number of these bits is then subtracted from the total cost. Those are "the bits you get back".

What I don't understand is what is the decoder going to do with these bits? How does knowing these bits reduces the overall transmission cost? Yes, the decoder is now aware of some bits that the encoder withdrew from a random-stream of bits, but how is this knowledge going to help in compressing the next observations?

[1]: Frey and Hinton. Efficient Stochastic Source Coding and an Application to a Bayesian Network Source Model

$\endgroup$
  • $\begingroup$ It's hard to understand if the paper [1] has any interesting contribution at all. Perhaps you can ask the question of the authors and report here--the whole setup seems somewhat artificial to me. $\endgroup$ – kodlu Aug 22 '15 at 10:01
1
$\begingroup$

I was wondering the same! But this is the answer I came up with (although I am not entirely sure if this reasoning is correct).

Since after the decoding process the decoder knows the distribution $Q(z|x)$ AND $P(x)$, he could then come up with an encoding that follows the distribution $P(z)/Q(z|x)$ and thus, represents each sample value $z$ with $-log_2(P(z)/Q(z|x))$ bits. Apparently, this representation of the sample values (and thus, of the model) would lead to maximal compression of the data values if the approximate variational posterior equals the actual posterior, thus $Q(z|x) = P(z|x)$.

$\endgroup$
0
$\begingroup$

I think that the key to understand bits-back coding is that sampling from the distribution $Q(z|x)$ is equal to decoding some "random bits" with $Q(z|x)$. And these random bits are the things that receiver can get automatically after receiving $z$ and $x$. Considering a sequential image transmission scenario, if we use the encoding results of previous image as the random seed for current image to sample from the $Q(z|x)$ in the sender part, these bits can then be subtracted from the total cost.

By the way, if the "random bits" used for sampling is not truly random there will be negative effects on your bits-back coding scheme.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.