Earlier quoted context omitted.
ah whoops, I'll fix it. ty!
Genuinely, from the bottom of my heart, thank you for writing without an AI.
everything worth writing, you should write yourself
Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
51–55 of 55 posts
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#52there is no such thing as "near lossless"
yes, your are right. what heading would you have taken here?
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#53Earlier quoted context omitted.
There is, after you define what you’re ready to loose and understand the lossy space. That’s how we came up with mobile cellphones, audio and video codecs etc. Literally powering all modern devices we use.
So then ... "lossy"
Yes of course. A compression algorithm which just stores the number of bits and decodes all bits as zeros is also lossy.
A floating-point compression algorithm which reconstructs the elements so they differ by at most one ULP[1] compared to the original value is, no surprise, also lossy.
Being able to communicate that an algorithm is closer to the latter than the former is useful, hence terms like "near-lossless".
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#54``` We evaluated several precision pairings across our internal retrieval benchmark suite. Scores are NDCG@10 averaged across the suite, scaled to 0–100. NDCG@10 (Normalized Discounted Cumulative Gain at rank 10) measures how well the top 10 results are ordered against the ideal ranking, rewarding relevant documents more when they appear higher, with 100 being a perfect ranking. The full-precision baseline averages 9…
>Saying "Near lossless" to mean 90% accurate retrieval of saved vectors is simply a lie. Lossy-ness is binary, not something you can paper over with getting close enough.
Lossy-ness is a binary, "near lossness" however is still valid (and is not the same as saying lossless).
How else, if not by comparing to "lossness" (whether with a more abstract qualitative term like "near" or with some distance or error measurement) do you report the level of fidelity to non-lossy results?
Depends on the context, but even in abstract for whatever domain, 90% sounds pretty close if we're talking about a linear level of degradation corresponding to each X% level.
In this case if this is calculated on a lossless baseline that's itself close to 90% - it's distance from it, it doesn't represent distance from some pure 100% perfect retrieval. So ~90% vs ~89% is very close to lossless capability.
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#55Earlier quoted context omitted.
> theres a big difference between 99% quality and 30%. sure > if you treat it in a binary way where everything short of 100 falls into one "lossy" bucket you lose all the practical differences that make one encoding much better than another. no; lossless is an inherently binary term. and I don't lose all the practical differences of better lossy encoders by understanding that; I'm not just going to start using mp3 96…
Lossless is objective for information theory. To get from the real world to digital world you need an analog to digital converter, this process is by definition lossy. We are interested in the real world, and information is pure but never represents exactly reality. Lossyness is baked into our problem statement here. Using terms like near lossless means we think we are very close to reality for what we’re trying to d…
ok
mp3 is near lossless