Talk:Vector quantization

"Some math"

An expression occurring in existential sentences. "For some x" is the same as " exists x." Unlike in everyday language, it is does not necessarily refer to a plurality of elements, and so might be more clearly represented in colloquial English as "for at least one." (Turkialjrees (talk) 16:44, 14 March 2015 (UTC)).

During some of my colleges I got some math what could be nice to be on this page. only I don't have enough mathimatical background to prove the used maths.

The Math

create set of prototypes = $w^{1},w^{2},\dots ,w^{K},w^{k}\in \mathbb {R} ^{n}$ the data = $\xi ^{1},\xi ^{2},\dots ,\xi ^{p},\xi ^{k}\in \mathbb {R} ^{n}$

by using the Squared_Euclidean_Distance we can determine the multidimention distance between a prototype and a data point. $L_{2}(a,b)=\sum _{i=1}^{d}|a_{i}-b_{i}|$ Based on this we can find the closest prototype to a given datapoint. $w^{i*}=argmin_{j}\{d[w^{j},\xi ]\}$ assign $\xi$ to prototype $i^{*}$

This way the winner takes it all and the closest prototype should be moved using: $w^{i^{*}}\xi \to w^{i^{*}}+\eta (\xi ^{t}-w^{i^{*}})$

where $\eta <1$ is the learning rate

Spidfire (talk) 15:29, 31 January 2013 (UTC)

clarity?

Damn. This article made me feel dumb. --NoPetrol 06:41, 24 Nov 2004 (UTC)

I have modified the article to give a clear explanation of what vector quantization is, together with some uses for it. It still needs tidying up and referencing Pog 21:46, 1 August 2007 (UTC)

also want to see pictures —Preceding unsigned comment added by 138.246.7.74 (talk) 13:50, 15 July 2010 (UTC)

Unclear sentence

"Find the quantization vector centroid with the smallest <distance-sensitivity>"

What does "<distance-sensitivity>" mean? Does it mean sensitivity? Or does it mean distance minus sensitivity? -Pgan002 00:17, 18 August 2007 (UTC)

I expanded it as distance minus sensitivity. But I think this is not a very good algorithm, and it may have been original research. So I added citation-needed because we need an established algorithm from e.g. some book. — Preceding unsigned comment added by 213.16.80.50 (talk) 14:42, 8 November 2016 (UTC)

Spam

Why the hell is there a picture of an aeroplane on this page? —Preceding unsigned comment added by Criffer (talk • contribs) 16:24, 11 October 2007 (UTC)

Definition

Is there a kind of agreed definition on this term? At least [1] attempts to define it. Should Wikipedia adopt this definition? Are there alternative definitions somewhere? Arkadi kagan (talk) 21:11, 25 January 2010 (UTC)

Another option from [2]:

A data compression technique in which a finite sequence of values is presented as resembling the template (from among the choices available to a given codebook) that minimizes a distortion measure.

Arkadi kagan (talk) 08:38, 28 January 2010 (UTC)

Use in data compression

"All possible combinations of the N-dimensional vector [y1,y2,...,yn] form the Gaurav."

What the hell is a Gaurav?

Secondly, even if there is a correct technical term for all possible combinations of an N-Dimensional vector, it is completely out of context in that particular article. It should be removed, or correct and given a context. —Preceding unsigned comment added by 198.151.130.16 (talk) 21:46, 1 April 2011 (UTC)

Where is a block diagram?

From the article: Block Diagram: A simple vector quantizer is shown below Huh? Where is it? Cuddlyable3 (talk) 09:15, 7 June 2011 (UTC)

Each cluster the same number of points?!

"It works by dividing a large set of points (vectors) into groups having approximately the same number of points closest to them."

This is not true, isn't it? E.g. clustering a 1-d normally distributioned data (10k samples) with k-means (6 clusters) results in groups with very different numbers of points assigned to each group (700 to 2400). I would not call this difference "approximately the same". Or am i missing something?

VERY approximate

From my limited experience, it seems most groups will have similar numbers, but a few groups (clusters) will have very few or very many elements assigned to it. So most clusters (maybe 60~80 %) will have a similar number of elements, but the remainder will have very few or very many elements. Hydradix (talk) 04:53, 13 October 2014 (UTC)

No mention of LBG or other methods

Article's "alternate training" method seems biased towards simulated annealing. No mention is made at all of the Linde–Buzo–Gray algorithm which is a fundamental starting point for most VQ implementations and is the most widely-cited paper in VQ work. No mention is made of PNN (Pair Nearest Neighbor) or other codebook generation methods either. --Trixter (talk) 19:49, 26 August 2013 (UTC)

Agreed! The LBG algorithm is fundamental for the topic, Vector Quantization. This, and other code-book generation methods, need to be referenced/linked. Although I have some experience with VQ, I am not an expert in VQ, so am not confident to update the page... Hydradix (talk) 07:43, 5 October 2014 (UTC)

update

I decided to be bold, and added in-page links to LBG and K-Means... I also added LBG to the References.... I tried/wanted to add Enhanced LBG to External References, but when I tired Wikipedia Preview the link would always fail (http://anale-informatica.tibiscus.ro/download/lucrari/2-1-02-balint.pdf) so ELBG was not referenced. — Preceding unsigned comment added by Hydradix (talk • contribs) 08:34, 5 October 2014 (UTC)

Article is too technical and abstract

I have no mathematical background. Despite my interest in signal processing, I didn't understand a word of the lede and used external information to add a sentence for the mortals among us. Once I gain a good understanding of the topic, I will update the article with more understandable information. --Holzklöppel (talk) 09:32, 11 October 2023 (UTC)