carbocation

I wonder whether LoRAs could be useful for U-Net training. Especially thinking of CNN-based U-Net models with pre-trained encoders (but randomly initialized decoders). At least, it seems possible that normal weight updates on the decoder and LoRA training on the encoder could improve efficiency.

Skipped a step: the model needs to be trained, which requires the patient symptoms as the target for weight updates. I think that you simply misread my original comment.

You said:

> Your comment misinterpreted this as “using the patients' symptoms and objective data” (when they only used objective data)

This represents an important misunderstanding of the methods of the paper. The model was trained using images (objective data) and the pain score (patients' symptoms). From the methods: "A convolutional neural network was trained to predict KOOS pain score for each knee using each X-ray image."

Also with respect to the author's claims, from the paper's abstract:

> Because algorithmic severity measures better capture underserved patients’ pain, and severity measures influence treatment decisions, algorithmic predictions could potentially redress disparities in access to treatments like arthroplasty.

You think I'm misinterpreting, but I still think that the paper is more important than you're giving credit.

Different poses? I don't think so, at least not with the current setup.

For example, I tried this with a dog (walking around the seated dog, taking photos as I did so). The dog turned her head while I was taking photos. The portion of the head that moved was not represented in the final output.

I'm not a radiologist so I could well be overinterpreting. However, if so, I am not sure that I am alone. This study published in Nature Medicine was hailed by radiologists as one of the "notable successes in using explainability methods to aid in the discovery of knowledge" [1].

Your sober assessment seems valuable, and would make for an interesting letter to the editor.

1 = https://www.thelancet.com/journals/landig/article/PIIS2589-7...

One related publication that I thought would have fit well into the comment about disparate pain treatment is this[1] from Ziad Obermeyer's group. They found that when they trained a model using patient pain and knee X-rays, much of the disparity in symptoms could be accounted for by findings from the X-rays themselves. It's a nice example of where using the patients' symptoms and objective data may actually outperform current medical standards, which fits with her participatory comments in the final paragraph.

1 = https://www.nature.com/articles/s41591-020-01192-7

Agreed. And not just the source, but a fully functional local demo! Runs great on my M1 pro using:

PYTORCH_ENABLE_MPS_FALLBACK=1 python3.10 demo.py --weights checkpoints/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth --device 'mps'

I played around with DUSt3R last night using an iPhone with different lenses (whether you consider this a different camera or not, I defer to you). It worked well. Note that the camera intrinsics aren't used here, so it makes sense that it would tolerate different lenses or cameras. I did not test wildly divergent lens types (e.g., a fisheye lens).

It seems unexpected that other people can edit a link to a Phind chat just by getting the URL. It means that if you share a URL with someone, they can change your results: https://www.phind.com/search?cache=k56i132ekpg43zdc7j5z1h1x

From clicking through the video and watching this summary, I didn't feel like this summary was very good. (No malice intended, just feedback.)

Very nice. Now let's see the external validation.

That is true for some diseases but not all. I don't think we know how to prevent dementia. But identifying biomarkers is useful, at the very least for selecting people for enrollment in future trials.

My workflow is similar to yours: python for deep learning and surface reconstruction. R for stats and plots.

I use go extensively for data preprocessing. Sounds weird but it works well for highly repetitive conversion tasks like DICOM parsing, converting EKGs to numpy, etc.

The link in this article to an analysis from the lab of highly enriched uranium from 1999 is a good read: https://inis.iaea.org/search/search.aspx?orig_q=RN:33057956

> Note, however, that SNPs like the one you pointed out are relatively infrequent.

If you mean that SNPs with small effect sizes don't always point to useful drug targets with big druggable effects, that is possible, but this remains an open question and is the subject of intensive research right now.

Illegal for health insurance companies. But not so for life insurance, or disability insurance or long-term care.

The effect size of common SNPs is not informative about the effect of drugging their related genes. For example, the common variants near HMG-CoA reductase have very small but significant (confidently nonzero) effects. Yet drugging HMG-CoA reductase can reduce LDL cholesterol by ~40-50% (statins).

See Figure 2 from this PNAS paper from J Craig Venter[1]. It caused quite a stir at the time, and not because people thought it was useful or wise.

1 = https://www.pnas.org/doi/full/10.1073/pnas.1711125114

Google scholar reports 22,000 papers with Windkessel in them; I've written a few. Pubmed reports over 800 with it in the title.

The concept of the Windkessel is cool. The notion that it hasn't been applied to the heart-aorta system is vexing. I feel like I must be misunderstanding.

Thanks for sharing, that is a very interesting coda to the story!

I wonder whether avo could have been useful here?[1] I mention it because it came up the last time we were talking about AVX operations in go.[2, 3]

1 = https://github.com/mmcloughlin/avo

2 = https://news.ycombinator.com/item?id=34465297

3 = https://www.reddit.com/r/golang/comments/10hmh07/how_to_use_...

Yes, I agree. OpenAI is wise to take this course, and the campaign won't be wrong to work around it by using an open model (in my opinion).

I use Waymo to get to work, and I prefer it over human drivers (even though it takes longer). It drives safely.

I can appreciate why companies would not want to be attached to politics in this way, or accused of influencing elections.

At the same time, as long as this person is consenting to this (which clearly seems to be the case), I don't think the developers did anything wrong (malum in se).

I suppose this is a use-case where people will turn to locally-run models.

This makes sense within the emerging literature about TLR7 signaling and lupus. There was a high profile paper in Nature in 2022 about lupus and TLR7: https://www.nature.com/articles/s41586-022-04642-z

It wasn't really obvious to me since I don't follow this closely, but NuScale is a company building nuclear powerplants.

Edit now that the link has changed, the new article makes this much more clear.

Bad wording on my part since I implied but didn't state the alternative. In my mind, the alternative scenario that would create a depressurization event would be a small hole or a hole that is not visible from the passenger compartment. I.e., a hole from which a passenger or attendant could not exit the plane.

Isn't this phrase generating ambiguity, in this circumstance, where a better description would not be ambiguous?

I know you're kidding, but we're trained to avoid making that mistake in assignment of cause of death (and in states where I've lived, they give clear instructions on the death paperwork to remind you not to make that kind of mistake)!

The nomenclature is interesting. In my mind, a "depressurization event" would imply that the biggest problem was the depressurization. In other words, that there was not a big gaping structural hole.

If there can be a big hole in the plane and we can call that a "depressurization event", I would like to advocate for changing how we describe this.