Location: Helsinki, Finland
Remote: Yes
Willing to relocate: Temporarily. Immigration will be easier in US (TN)/Canada /Australia. Not an EU national (yet).
Technologies: Deep learning, though I'm seeking something very specific, see below.
Résumé/CV: https://sspilsbury.com/hire
Email: s@polysquare.org
I'm a 2nd year Machine Learning PhD student looking for an internship to work on something related to my thesis. The thesis is on compositional generalization problems in deep learning, right now with a particular focus on NLP and interactive environments. Asides from that, I have a lot of background in a previous life doing research/graphics/UI/web/backend/ETL stuff at various other places.
Some open research problems which might be relevant to what you are working on:
- Much improved sample efficiency for language-interactive robots or agents (don't need to train every single combination of different aspects)
- Compositional path synthesis
- Compositional particle or protein synthesis (eg, composing two separate subgraphs into a single graph by way of following some learned rules)
- Understanding why transformers can and can't do different kinds of compositional generalization
- Improved sample efficiency and generalization on many different language tasks in different language domains (for example, grammatical error correction, machine translation, semantic parsing, entity and relationship extraction).
One hard requirement is that we must be able to publish the research results of my internship such that it can be included in my thesis.
The key term here is "reside". We'll get back to this in a bit.
> Professional athletes pay taxes to every state where they play a game.
You are confusing tax-at-source with residence based taxation. Which is fair enough because many countries try to do both.
For the purposes of these next few paragraphs, lets say that country X is your country of residence for tax purposes and country Y is where you happen to be performing some work temporarily.
Source-based taxation means that if you are paid from a source in country X (or the income you make has a sufficiently strong connection to country X), then you pay tax to country X on that income. Typically the tax is withheld by the payor as opposed to the payee needing to file a tax return. This is what applies to your example of professional sportspeople. There's a very clear link - you play the sport in country X, you get paid by the competition in country X and you pay some tax to country X. This income might also be taxable in your country of residence, but that's a different issue where tax treaties and paid-foreign-tax deductions come into play.
With remote work for a foreign employer getting paid into a foreign bank account where the work doesn't have much connection to the country, the link is less clear. For short stays, many countries will not consider this to be locally sourced income. See this example from the Australian Tax Office which answers this very question (https://www.ato.gov.au/General/COVID-19/Support-for-individu...). Note that in this case, you are still only paying tax on locally sourced income and not worldwide income.
Then there is the question of "tax residence". Different countries have different rules about this and residence is not "exclusive" (so you can be multiple-resident if you're unfortunate in how you set up your affairs). Tax residence in most places happens after a fixed period of stay in the country (typically 183 days) and/or if you have "residence ties" to that country. "residence ties" is typically a multi-factor balancing test, which includes things like owning real estate, supporting a spouse and dependents who continue to live in that country, having a fixed address in that country, having your essential social connections (club memberships, service subscriptions etc) run out of that country, nationality etc. Short non-successive trips to a country don't usually create residence ties. Most countries follow this model (https://en.wikipedia.org/wiki/International_taxation#Source_...). Then there's Eritrea, Hungary, Myanmar, Tajikistan, the United States for which citizenship automatically counts as tax residence, but usually there's offsetting procedures and tax treaties to avoid double taxation even if there might be double filing.
Then there's the "working on a tourist visa" question. This is unfortunately a significantly more murky area, especially when it comes to remote work. A good rule of thumb is that coming to a country for the purpose of remote work and not for the purposes permitted under a tourist visa or visa waiver is possibly over the line, but replying to some emails, attending a few meetings via call, fixing a bug here or there while you're mainly on holiday is likely to be fine. Of course the safest bet is not to do any work at all, but if we were to apply some common sense here, it would be absurd that you get an entry ban for the apparent crime of replying to an urgent email from your boss or colleague while on holiday.
Then there's the question of payroll taxes and unemployment/health insurance contributions to be paid by your employer. I don't know very much about this, but I think a starting point in the analysis would have to be whether your foreign corporate employer is subject to any sort of personal or tax jurisdiction by the country that you're in at all. Maybe someone else will fill in on this one.
> now we have even less freedom of speech inside the country.
Did anybody read this part?
I see a lot of comments along the lines of "well we're at war with you" or "you'll get your access back once Ukraine stops being bombed", with the justification that anyone living in Russia has the duty to "rise up" and stop the war locally. To be these responses seem thoroughly un-nuanced and flippant. Its absolutely true that to stop the war, you need to get rid of Putin, but its not clear that there's a connection between dropping clients and getting rid of Putin. In fact this is probably counterproductive and strengthen's Putin's grip on the population even further.
The usual mechanism given is "if we annoy the people enough, they'll rise up on the streets and demand change". First, this assumes that if you annoy people, they'll transfer their anger to the regime. This is only true for a subgroup of affected people; the rest will just be angry at the sanctioning party and direct their anger through the regime into even stronger pro-war support. But for the people who might be motivated to protest, its not clear whether or not their actions will be effective. Because we're talking about a dictatorship, individual action carries both high personal risk and is unlikely to be very effective. There are exceptions to this rule, for example, Marina Ovsyannikova, but ordinary people going to the streets and holding up signs is only likely to end up with more people in jail (and potentially more people conscripted into the army). There is also the objection that mass-protest will make a difference, which, perhaps it will, but that requires co-ordination and a critical mass.
The second thing to say here (relevant to what I quoted from OP) is that Mailchimp is infrastructure which is in use by the local resistance. By blanket cutting-off Russia, you're also hindering the local resistance as well. If one wants to take a principled stand, its probably fine to cut off regular businesses that have nothing to do with it (with the caveat that you might just make them angry at you), but throwing out the baby with the bathwater and cutting off the local resistance is just completely counterproductive. Due to the inability to make payments outside the country, their only options for replacement infrastructure are local, which are more likely to be under state influence and censorship. So now their only options are to shut down and stop organizing the resistance to the regime, or engage in self-censorship, which might as well just be shutting down.
From the looks of it, they're leveraging a fork of a library I ported from Compiz a few years ago (called libanimation). I'm quite happy to see this, as I thought it was a dead end and wouldn't get any users, but it looks like its been very useful to this project and I'd be interested to see where the development is continuing so that I can help them out.
Wobbly windows is such a truly excellent feature; I really don’t understand why more platforms (most significantly Windows and macOS) don’t go with it, even if more restrained (Zorin’s jelly looks a bit too loose for my liking). They make the entire experience much more tactile and make windows all round much nicer to work with.
(For my part, I now use tiling window managers, where wobbly windows won’t work well except for floating windows. But on all-floating window managers—)
> One annoyance is that the graphs don't render if they're slightly off-screen (and they render slowly/fade-in or something) so it can be somewhat annoying to scroll back/forth between graphic examples.
Yeah. This was an engineering tradeoff - there's a limit on the number of WebGL contexts (15 for Firefox, for instance) that can be running at any time so I had to turn off rendering of the visualizations when offscreen.
If anyone knows a better way to approach this, please let me know!
> There are a few grammatical errors (missing word, missing comma) that are forgivable but I find that any imprecision in language is a big detriment for a student trying to understand an explanation. I tend to get hung up on those since, not knowing the material, I can't resolve ambiguities with context. Since subtleties are important in math it would be worthwhile to go over the explanations again with a fine-toothed comb to ensure there are no typos or amphiboly.
Yeah, this is a pet peeve of mine for sure. I am not very precise when it comes to written prose.
I was thinking of airtasking out a grammar check, since I'm pretty awful at catching my own mistakes here. I might just do that.
> The visuals aren't always clear. For example, volumes don't show edges so it's hard to make out that you're seeing a volume rather than an irregular area (and which volume you're seeing). Also I think a caption to each image could be useful, to describe how the shape demonstrates the concept. Something like "this visual represents a shear transformation of the matrix described above", just to be really explicit and clear.
> This seems interesting, but I feel like it would be way more useful if you could drag to rotate however you want, instead of waiting for the animation to get sort of close to the viewpoint you're interested in.
Great feedback! I had been thinking of doing something like this but I was waiting on actual readers to mention it. I'll see if I can get to it this weekend.
> I really like the visual representation of elementary row operations, but the intersection of the three planes is off their edges, so isn't actually shown.
Yeah, good point. In the last visualization we have a vector pointing to the intersection, perhaps a similar sort of thing can be leveraged there.
> I found a bug (Chrome 65.0.3325.181, Windows): The scroll position is kept when you switch sections, so you have to manually scroll up to the top.
Thanks for the report! I think someone else mentioned this already and I filed a ticket on the github repository.
Some papers I'm a first author of include:
- Compositional Generalization in Grounded Language Learning via Induced Model Sparsity (https://aclanthology.org/2022.naacl-srw.19/)
- Meta-learning from demonstrations improves Compositional Generalization (https://openreview.net/forum?id=hb3Et9tJSC9).
Some open research problems which might be relevant to what you are working on:
- Much improved sample efficiency for language-interactive robots or agents (don't need to train every single combination of different aspects)
- Compositional path synthesis
- Compositional particle or protein synthesis (eg, composing two separate subgraphs into a single graph by way of following some learned rules)
- Understanding why transformers can and can't do different kinds of compositional generalization
- Improved sample efficiency and generalization on many different language tasks in different language domains (for example, grammatical error correction, machine translation, semantic parsing, entity and relationship extraction).
One hard requirement is that we must be able to publish the research results of my internship such that it can be included in my thesis.