My (possibly wrong) TLDR: TransMLA is a method to "compress" an already trained GQA model, with the additional option to further fine tune it. Shall make inference faster.
It is not a method to compress a Grouped-Query Attention model, but to expand it into an equivalent Multi-head Latent Attention model with the same key-value cache size but larger effective key/value vectors and a correspondingly larger number of trainable parameters. With additional training, you can then obtain a better model that only uses a little bit more memory.
I am working on the sunflower plant density estimation problem. The goal is to be able to estimate the germination rate as early as possible. Farmers benefit from such information, because:
- there are lots of expenses still to be made (fertilizer, pesticide, salaries), which may not be worth it if germination is under certain threshold
- if detected early, there is still time to plant another grain or to fill up the missing plants (requires precision seeders and seeding maps)
- is a very good proxy for yield estimation (farmers often trade futures even before they have harvested)
For the purpose I have created a dataset (a collaboration between my employer and Sofia University) and published it in order to enable scientific collaboration with other interested parties. Still working on the dataset annotations.
Interesting, I'm also involved in a project to do yield prediction, but with a ground-vehicle with camera's on top to drive between strawberry and blueberry plants.
Yield prediction is huge indeed, because overshooting your prediction means seller stuff for a lower price. Undershooting means paying for someone's product to make up for the difference. Probably there's quite a bit of matchmaking in between those under and overshooters and someone making a good buck out of that too.
> Undershooting means paying for someone's product to make up for the difference
Indeed. Making up the difference can easily eat most of the farmer's profits. I guess it is even more pronounced for berries when compared to grains, because they cannot be stored for so long.
Hey, this is interesting. I used to work on a somewhat similar problem. Our problem was more general, but one usecase is to predict the number of interactions between flowers and pollinators, given some initial counts. As these initial counts are obtained manually (by going to the fields, taking pictures and count, like number of bees within a frame), those count numbers are likely to be lower the the actual numbers. We addressed this under-counted issue using low-rankness and Poisson mixture model. Take a look if you're interested: https://ieeexplore.ieee.org/document/10888717
It is possible. However, getting accurate yield data requires a "smart" harvester that can produce yield maps. Many modern harvesters are equipped with GPS and various sensors, so it is possible. However, farmers are really slow to replace old equipment if it works fine. I guess there are some retrofit solutions for yield mapping, but I haven't investigated their affordability and penetration into the (EU) farming landscape yet. Additionally, there are other interesting parameters apart from the harvested quantity that can be captured (e.g. the quality of the grain itself, such as size, composition, humidity etc).
I am not sure that I understand your question correctly, but given more precise sunflower density estimation, the farmer has three options:
1. Plow the field and seed again (same or different variety or grain). This is a very crude measure, but it is sometimes the right thing to do, because as I said most of the expenses have not been realized yet (fertilizer, pesticide, fuel, payroll, paying rent for the land). It is also a time critical decision, because the window of opportunity for plowing and reseeding is not very wide.
2. Accept the lower yield if it is within a reasonable margin (e.g. comparable to the expenses to plow and reseed).
3. Do partial reseeding over the existing plants (without plowing). This is an emerging strategy with the proliferation of smart seeders, but it requires a precise seeding map to be created beforehand (i.e. based on the density estimate). As an advantage, you spare the expenses for seeds and plowing, however there is some disadvantage as well, due to the different rate of development of the newly seeded plants. Farmers usually need plants to be ready for harvest at the same time, otherwise the quality of the grains suffers and hence the selling price is lower.
In addition to these points, having precise density information after germination helps with the identification of problems, such as seeder malfunction (e.g. nozzles getting clogged), seed quality and meteo data (e.g. too much rain, low temperatures etc).
In the EU they use fertilizers and pesticides, but rarely use irrigation. However, pesticides are usually applied over the lifecycle of the plants. For fertilizer, there is some value to apply it on time, because it tends to migrate with water and applying now vs a month ago is not equivalent.
Indeed. I am just a bit concerned that the subjects were to return to their "normal" diet after the experiment, which probably was the reason to get this condition in the first place. In a next study it would be interesting to look closer into the functioning of related organs, such as liver and gallbladder (vesica fellea). E.g. did subjects have any biliary sand issues. Correlation with potable water used by the subjects would also be interesting, especially things like mineral composition, pH, organic contaminants (e.g. microplastics) as well as microbial content.
It’s amino acids, fats, and sugar in a ratio that mirrors the elements of a standard diet.
It’s not a supplement. It has to replace your diet completely for 2 weeks, no exceptions.
By consuming amino acids instead of proteins (which are composed of amino acids) and omitting fiber it starves skips straight to absorption and starves the bacterial overgrowth.
It’s an old concept. The unique part of this one is that it supposedly tastes okay.
Yeah, the taste matters if you are going to eat it for 2 weeks :). You will probably still hate it after the first week, but it is just one week of misery instead of two. I did some fasting in the past, leaving only water, tea (no sugar) and peeled apples in my diet. Had to do it for 2 weeks, but only made it to the 11th day. The issue with that one was the low calories, which doesn't correspond well with active work (I did not plan properly to take the second week off). Nevertheless, it helped solve the colitis issue, with which I'd been struggling for a few months. An interesting observation was that up to say day 4 or 5 you actually have more energy and sleep less. But after that it was a bit of a struggle to retain the same levels of energy and be productive at work.
The secret sauce, so to say, is that it's not a supplement, it's all you eat for the duration of the treatment. As someone who was on a liquid only diet for six weeks following a surgery, I can't imagine this being anything short of absolutely miserable.
It is not better than GPS. It is better than traditional inertial navigation systems (INS). But the accuracy is sub 500m for a good portion of operations vs multi-km resolution for traditional inertial systems.
Yep, the title is completely wrong. The actual article says ".. INS", not GPS. It can't compare with GPS to begin with. From the article: "the best final positioning accuracy we achieve is 22m". GPS can be accurate to a centimeter level, even inaccurate (no other reference) GPS is at least accurate to about ten meters.
For plenty of military applications, 90m accuracy is a valuable fallback in a GPS denied environment. It's probably not nice for targetting purposes. But for general orientiering and the question 'are friendlies in this area' it's a lot better than nothing.
In my experience, military aircrafts are visible in flightradar most of the time. However, if they are flying in a group (e.g. choppers), usually only one of them is visible.
