On Tue, May 25, 2021, 9:20 AM stefan.reich.maker.of.eye via AGI <
[email protected]> wrote:

> On Monday, May 24, 2021, at 10:00 PM, Matt Mahoney wrote:
>
> To code a word prediction, you have to calculate the probability
> distribution over the entire vocabulary.
>
> Wait. You're saying I have to scan my entire vocabulary when I predict a
> word someone says? I don't think that's what anyone does.
>

In a neural network, the n probabilities can be calculated in parallel in
one step. To arithmetic code the result you need to assign a fraction of
the range space to each word in proportion to their probabilities. This
requires at a minimum adding them up, which can be made parallel up to log
n steps using an adder tree. That's the same as predicting log n bits
serially, which is what my code does.


------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tf856e4082d9ea09a-M74aeddab10e25bb35bd6e1a9
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to