Comment by PaulDavisThe1st | Hacker News Reader

Comment by PaulDavisThe1st | Hacker News Reader

PaulDavisThe1st

A nice introduction. However, I feel that many such introductions, it skips too quickly over the question of feature selection.

This is the step that introduces, rather subtly and sometimes almost ignorably, human judgement into what otherwise feels like a very automated and "it's just math" system.

Let's take audio. What features will you select from it to make your N-dimensional vector? An easy answer might appear to be "as many as possible". However, your first problem is that you may not have access to feature characterization even for just the features you can trivially name. Your second problem is that without deep knowledge of the domain, you may not even be aware of potential features that you should probably be using. Your third problem is that even with deep knowledge of the domain, you may still be unaware of potential features that you probably should be using.

A practical example may help. Perhaps you're a fan of Reichian phase-shifting minimalist music. You (indirectly) interrogate a vector database to find music similar to a canonical piece from this genre (say, Piano Phase). The database uses many audio & music related features, including but not limited to: dominant frequency, note onset intervals, volume, frequency-distribution-based timbral characteristics, apparent root notes and scales etc etc etc.

Unfortunately for you, the feature set in the database does not include "note-to-note intervals over time is constant". Your query will manage to find things that are timbrally, harmonically, melodically and rythmically similar, but it will be by sheer good luck that any of them share the key characteristic: steady changes in the relative phase of two different (or the same) melodic lines.

This is just one example. It's not hard to cook up equivalents for visual, textual, numerical ... any data at all.

Now of course, none of this makes vector databases and feature classification useless. It does mean that when you do or do not find match patterns in a given data set, one of the early questions you need to ask yourself is whether the feature set has any strong guarantee of being complete, and if not, how it may need to be expanded.

Vector databases are oriented around search and retrieval. The usual approach of generating vectors is to fine tune a large pretrained model and extract the inner representations. Because the dataset contains successful queries and retrieval results, all you need to do is optimize a loss function (usually contrastively or approximated versions of common ranking functions) using raw inputs on the similarity objective supported by the vector database. For common modalities like tabular/text/image/audio data, there is basically no human judgement involved in feature selection - just apply attention.

Note: current state of the art text to vector models like E5-Mistral (note: very different from original E5) don’t even require human curation in the dataset

This is also just eliding the work that humans still have to do.

I give you an audio file (let's say just regular old PCM wav format). You cannot do anything with this without making some decisions about what happens next to the data. For audio, you're very minimally faced with the question of whether to do a transform into the frequency domain. If you don't do that, there's a ton of feature classification that can never be done. No audio to vector model can make that sort of decision for itself - humans have to make the possible.

Raw inputs are suitable for some things, but essentially E5 is just one that has already had a large number of assumptions built into it that happen to give pretty good results. Nevertheless, were you interested for some weird reason in a very strange metric of text similarity, nothing prebuilt, even E5, is going to give that to you. Let's look at what E5 does:

> The primary purpose of embedding models is to convert discrete symbols, such as words, into continuous-valued vectors. These vectors are designed in such a way that similar words or entities have vectors that are close to each other in the vector space, reflecting their semantic similarity.

This is great, but useful for only a particular type of textual similarity consideration.

And oh, what's this:

> However, the innovation doesn’t stop there. To further elevate the model’s performance, supervised fine-tuning was introduced. This involved training the E5 embeddings with labeled data, effectively incorporating human knowledge into the learning process. The outcome was a consistent improvement in performance, making E5 a promising approach for advancing the field of embeddings and natural language understanding.

Hmmm ....

Anyway, my point still stands: choosing how to transform raw data into "features" is a human activity, even if the actual transformation itself is automated.

I agree with your point at the highest (pretrained model architect) level, but tokenization/encoding things into the frequency domain are decisions that typically aren’t made (or thought of) by the model consumer. They’re also not strictly theoretically necessary and are artifacts of current compute limitations. Btw E5 != E5 Mistral, the latter achieves SOTA performance without any labeled data - all you need is a prompt to generate synthetic data for your particular similarity metric.

> Unlike existing methods that often depend on multi-stage intermediate pre-training with billions of weakly-supervised text pairs, followed by fine-tuning with a few labeled datasets, our method does not require building complex training pipelines or relying on manually collected datasets… We leverage proprietary LLMs to generate diverse synthetic data for hundreds of thousands of text embedding tasks across nearly 100 languages.

It’s true that ultimately there’s a judgement call (what does “distance” mean?), but I think the original post far overcomplicates what’s standard practice today.

Sorry, I just not believe this generalizes in any meaningful sense for arbitrary data.

You cannot determine frequencies from audio PCM data. If you want to build a vector database of audio, with frequency/frequencies as one of the features, at the very least you will have to arrange for a transform to the frequency domain. Unless you claim that a system is somehow capable of discovering fourier's theorem and implementing the transform for itself, this is a necessary precursor to any system being able to embed using a vector that includes frequency considerations.

But ... that's a human decision because humans think that frequencies are important to their experience of music. A person who totally deaf, however, and thus has extremely limited frequency perception, can (often) still detect rythmic structure due to bone conduction. Such a person who was interested in similarity analysis of audio would have no reason to perform a domain transform, and would be more interested in timing correlations that probably could be fully automated into various models as long as someone remembers to ensure that the system is time-aware which is, again, just another particular human judgement regarding what aspects of the audio have significance.

I just read the E5 Mistral paper. I don't see anything that contradicts my point, which wasn't about the need for human labelling, but about the need for human identification of significant features. In the case of text, because of the way languages evolve, we know that a semantic-free character-based analysis will likely bump into lots of interesting syntactic and semantic features. Doing that for arbitrary data (images, sound, air pressure, temperature) lacks any such pre-existing reason to treat the data in any particular way.

Consider, for example, if the "true meaning" of text was encoded in a somewhat Kaballah-esque type scheme, in which far distance words and even phonemes create tangled loops of reference and meaning. Even a system like E5 Mistral isn't going to find that, because that's absolutely not how we consider language to work, and thus that's not part of the fundamentals of how even E5 Mistral operates.

The above is a common anthropocentric take that has been repeatedly disproven by the last decade of deep learning research: http://www.incompleteideas.net/IncIdeas/BitterLesson.html

Understanding audio with inputs in the frequency domain isn’t required for understanding frequencies in audio.

A large enough system with sufficient training data would definitely be able to come up with a Fourier transform (or something resembling one), if encoding it helped the loss go down.

> In computer vision, there has been a similar pattern. Early methods conceived of vision as searching for edges, or generalized cylinders, or in terms of SIFT features. But today all this is discarded. Modern deep-learning neural networks use only the notions of convolution and certain kinds of invariances, and perform much better.

Today’s diffusion models learn representations from raw pixels, without even the concept of convolutions.

Ditto for language - as long as the architecture is 1) capable of modeling long range dependencies and 2) can be scaled reasonably, whether you pass in tokens, individual characters, or raw ASCII bytes is irrelevant. Character based models perform just as well (or better than) token/word level models at a given parameter count/training corpus size - the main reason they aren’t common (yet) is due to memory limitations, not anything fundamental.

For further reading, I’d recommend literature on transformer circuits for learning arithmetic without axioms: https://www.lesswrong.com/posts/CJsxd8ofLjGFxkmAP/explaining...

Great comment about you don't know what you don't know. Also wonderful to see another Steve Reich aficionado. What are some similar pieces to music for 18 Musicians which I adore?

There's nothing quite like Mf18M.

However, I do love several of the pieces by Marc Mellits, check out "Real Quiet plays the music of Marc Mellits". All rather short, but some totally lovely.

You may enjoy an 80's album by Andrew Poppy, in particular the track "The Object is a Hungry Wolf".

Do be sure to check "Shift" by Chris Hughes, which are re-instrumentations of several early Reich pieces including Piano Phase.

It's quite different in many ways from Mf18M, and you may have heard it, but I adore "The Chairman Dances" by John Adams. It couples the "chugging along" of Mf18M with much less minimalist melodic and harmonic structures. YMMV.

I'm sure I am going to remember a bunch more after I send this, so check back here in a couple of days, and I'll reply again with any extras.

Do you have any "similars" ?

Reply on news.ycombinator.com