jphoward

Isn't Help to Buy only for first time buyers, so it biases against Buy to Lets?

I think links like this are really helpful. Too offer I suffer from the (common on HN) mindset of, "I know nothing about this field, but I'm intelligent enough to know this is stupid". Posts like this are not only interesting, but give me a small dose of humility.

Exactly, the prototypical example has implemented it is one without democratic elections and so has no risk of being voted out of office.

I am all for writers taking a balanced perspective, especially when it comes to something as subjective and controversial as programming languages.

However, I do find it a bit frustrating when someone writes a 10 page article and concludes with "you can use either". It's actually not helpful to the reader, because they do have differences, which you've talked about extensively in the preceding 10 pages. If they were identical, you wouldn't need the article, and OBVIOUSLY you can use either.

Indeed, in the post above he made valid points such as Go may be thought to prioritise speed of development over speed to execution when compared to Rust.

It's absolutely fine to state an opinion at the end of an article which is clearly on a subjective passage.

I can absolutely see Netflix making a smash hit docudrama about this trial. I'm not being facetious; miscarriages of justice really do capture the public's imagination, if they're able to hear the story told in an interesting way. Add in to the fact that this is a political thriller, and it's a recipe for success.

And you 100% know Assange will have 0 qualms with selling the rights, and extensive phone calls from prison.

So I've actually tried to implement Eulerian video magnification myself, albeit using Python, because I believe when I first looked into this, ~5 year ago, the code was MatLab only.

I was really excited about it, because, as a cardiologist, I immediately thought we could use this to identify irregular heart beats in the patient waiting room, for example. This was pre Apple watch etc. I was thinking people could briefly sit in the chair in front of the camera if they are willing to be recorded, as a quick screening whilst they wait for the doc.

Unfortunately, I found it incredibly difficult to reproduce their results, even using similar data. I remember videoing my foot as I have a relatively prominent posterior tibial pulse, which in good light I can actually make out visually (but not in the video I took, as intended). The Eulerian magnification didn't seem to do anything.

Has anyone else had better luck? I did wonder if maybe I had to use specific video capture equipment, certain frame rates or filters, but if so that wasn't clear from the documentation?

It's been 8 years since this paper came out, and yet I still haven't seen it being used in the 'real world', so I am slightly suspicious that I am not alone in my experiences.

The use of gradient accumulation here is interesting, and I'd argue maybe not optimal.

He uses a batch size of 8 (because his GPU only has enough RAM to allow 8 CT scans to be tracked through the network at once). He acknowledges that this is smaller than he would like, and so he uses gradient accumulation so that he only adjusts the weights of the network every 8th forward pass. Effectively, he averages the results over 8 batches.

The thing about batch size is it's a trade off. It turns out, even if your GPU could support you putting your entire dataset through at once, this isn't a good idea. The explanation for this can be simplified as "if you give it the same data every time, it'll give you the same answer every time". This may sound fine, but actually a bit of randomness is very helpful to get the network outs of local minima, i.e. local ruts. This is why we do "mini batch" stochastic gradient descent. Now his dataset is only 200something patients and 300something scans, meaning he only gets 4-6 batches out of his dataset for every epoch (every cycle). I would expect (though I have no proof) that using a lower batch size might increase stochasticity here.

The problem is, smaller batch sizes also have their problems. The main reason for this is due to the ubiquity of a layer in CNNs called batch normalisation (BatchNorm) layers. These layers calculate on the fly (during training) what the mean and standard deviation across every feature in the current batch is, and rescale them so together they mean of 0 and a standard deviation of 1. The problem is, when your batch size is very small, your means and standard deviations might be wildly inaccurate (e.g. if your batch size is 4, it wouldn't be that surprising if every sample in the batch is a COVID patient, or every sample is normal). The batchnorm layers will 'remap' those images onto a normal distribution, and make a batch of 4 COVID patients look more normal, and nice versa. For this reason, we try not to train networks with batch sizes below 10, because when BatchNorm is involved, things tend to start behaving badly. Unfortunately, BatchNorm doesn't give a damn about gradient accumulation, because they are still different batches. Gradient accumulation does not get around the biggest problem of small batch sizes, and in my experience very rarely helps. Instead, you're better of using a different normalisation layer.

I absolutely love this! But during busy times many lines have trains separated by <= 2 stations (often < 3 minutes between trains on Victoria line), so one would expect most of the lights to be on. How does it decide which trains to draw?

One thing that surprises me (and may be the source for some of his frustrations) is how much for a power user he uses the mouse. This isn't a criticism, just an observation.

He, fairly, criticises how long it takes for the filename to show up when he hovers over an icon on the task bar. However, I think the vast majority of people at his level would use Alt+Tab which immediately brings up not only larger icons for active applications to swap between, but the filename above.

Link appears broken, but hopefully this is it: https://github.com/ThomasHabets/goircd

Is it fair to say JAX is what Google made when they looked at Pytorch/Autograd and thought, "oh damn, that's what we should have done?".

If so, is this the beginning of the end of Tensorflow? I know Tensorflow is still top for production, but it is certainly rapidly losing followings in the research field, and Pytorch and now starting to focus on deployment as they know this is their weakness.

This is not at all meant to be a criticism, but more a question: is taking a 20% performance hit, as the author writes, really worth it for a theoretical improvement in the lifespan of some integrated circuits? Are ICs actually how modern laptops break?

What do you think the odds are in him realising this benefit? Assuming he bins the laptop after, say, 5 years? (he's rich enough to buy a £1.3k minimum laptop, assuming it wasn't second hand)

It relies on several assumptions: 1) His laptop dies before he bins it in 5 years' time 2) It died becuase of an integrated circuit (the reason he cites) - in my experience old computers tend to die of things like exploded motherboard capacitors rather than ICs. Maybe the death of these are accelerated by turbo boost, too? 3) The integrated circuit's death was hastened by a meaningful amount (several months at least; a dodgy thing dying a couple of days/weeks later doesn't save you from buying a new laptop).

Is taking a 20% performance hit on your laptop really worth this? I'm interested. I can't imagine it would be for the vast majority of people.

Diseases are rivalled only by war in bringing out the best in human creativity and ingenuity.

I'm surprised and pleased about how similar the toys are to those from when I was little. Lots of train sets and cars.

I suspected to see a series of iPads and games consoles. I flicked through 20 or so and barely a battery-powered toy in sight, never mind a computer.

I don't know why it should make me happy though... maybe I'm worried modern children's lives are too complicated and just idolise my own, without any evidence for that. Getting old, I suppose.

Ignoring the fact that the title of this post is misleading and also not used by the linked article, I was interested to read that MaxNorm was _so_ effective. In my experience it is rarely used in the state-of-the-art ConvNets, though maybe that is because they are trained on large ImageNet datasets where overfitting is less of an issue? Weight decay/L2 norm seems almost ubiquitous is comparison.

Have other HN readers found MaxNorm to be that useful? Am I missing out?

That sounds incredible to me as using pandas without falling into the traps of chained indexing seems so much harder than 90% of Python.

For anyone who knows intermediate python and numpy and wants to learn how neural networks (CNNs, RNNs) work through simple pythonic examples, Chollet’s book is excellent. I’ve bought two copies as I can’t bear lending mine out to my students. He’s the creator of the Keras library and it’s beautiful in its symplicity.

Yep but the strikes (supported by a poll of members) fizzled out and then a second round if industrial action was decided against by the BMA (without a further vote).

And possibly the most unpopular union in the UK, too. I've parted ways with them, as have most of my friends. It's not because I disagree with their decisions, as much as they are completely impotent.

I do empathise; the BMA are stuck between a rock and a hard place, in that their members want better work conditions, pay rises that actually match inflation (shock horror), but the government just say no.

The difference between the other unions and the BMA, though, is the BMA go back to their members and say "the government said no, sorry", whilst the others say "so be it", and the strikes commence.

I'm not quite sure how I'd improve the situation if I was a BMA representative; it's complex. However, there is literally no reason currently to pay the membership fee.

Time and time again we see epidemiological studies posting findings that are clearly best explained by correlation than causation. For example, think of the people you've met with the worst teeth. They will almost certainly, on average, have worse overall health, lower socioeconomic status, lower intelligence, lower paid jobs, lower educational attainment, worse diet and so on. Surely these are just as likely, if not more, to be on the causal pathway for Alzheimers? And that's completely ignoring the fact that people with early Alzheimers which has not been diagnosed are much less likely to remember to keep up with oral hygiene rules, so the risk for reverse causation is HUGE.

Now I know you're going to argue that the epidemiological studies account for this using multiple regression and so on, but the fact is time and time again it has been demonstrated that this just cannot be done. Indeed, vitamin D in practically every epidemiological study to date has shown to reduce the risk of cancer, cardiovascular mortality, tuberculosis infect. And yet, just this month, someone FINALLY did a randomised trial of vitamin D in TWENTY SIX THOUSAND people (the VITAL trial) and found it made no difference at all (in fact, cardiovascular death was slightly but non-significantly more common in the vitamin D group).

Unfortunately, people's careers and publications rely on this stuff, so it'll keep coming. Please remain skeptical, hacker news.

What I don't understand is if you can train a final layer on the reservoir's random representation, why is this better than just training the final layer on your data directly? I assume the answer has something to do with dimensionality reduction?

Actually these trials were about expanding Mitraclip's indications.

The old EVEREST trial resulted in it being licensed for degenerative valve disease (damaged valves), but here they wanted a license for treating otherwise healthy valves which are leaky due to being stretched open by a dilated ventricle in heart failure (functional valve disease).

So this is currently fairly controversial in the cardiology world. This trial (COAPT) is tremendously positive, but only months earlier a very similar trial run from France, MITRA-FR, was dead negative (neutral).

Some of the authors of COAPT are trying to explain this by slightly worse valvular parameters and larger heart chambers. MITRA-FR also had a slightly higher complication rate.

However, my personal feeling is we need a third trial.

Agreed, the opening paragraphs talk about "overthrowing C", the release being "massive", lots of "I" and "my", and also the statement you highlighted.

I'm sure this an excellent language, but it sounds like it's written by a megalomaniac.

I'm sure it's not intended that way, but I would advise the author to tone it down a bit so people don't take it the wrong way.

The wonderful thing about Krita isn't it's features (although they are excellent); it's the fact it's the only graphics program with modern capabilities that someone could transition into from Microsoft Paint without needing to load up a series of YouTube tutorials.

I commend the Krita team for not merely copying Photoshop/GIMP, but actually writing a program an idiot like me can pick up and actually be productive with.

In the field of medicine this is standard, at least in the UK. Many medical specialties expect you to have a PhD to get a good senior post in a teaching hospital. To get a good PhD grant you are expected to have a pedigree of publishing research, and the classic way to do that is join a research group on your days off/evenings/weekends.

I’m lucky enough to be on a good PhD programme so my 9-5 is now academia, but in my late 20s a typical weekend and weeknight would involve a couple of hours of academic writing which I would be doing “for free”. I’d be first or second author on the papers, but I’d be publishing under the institution of the group I’d tagged on to.

...you wrote a weather forecast?

I wonder who the target market is, though? The iMac Pro can cost £5000, so I assume more performance and hence price than that? The only users who I can imagine willing to justify paying such a premium are those who feel they must remain at the cutting edge of performance and are willing to pay for it, but surely they will have left the Apple ecosystem by then?

It's as if Apple are saying "we're making a computer for people who demand peak performance, but are willing to use cheap outdated hardware until we get round to it."

It’s striking how little humility there is in the post. The majority of it sounds like a sales pitch, and I suppose, who can blame them if it is one in disguise?

How valid is studying Mensa to get a representative cross section of intelligent people? I would wager many on this site are intelligent enough to qualify for Mensa, but why would they join? Most will be sufficiently surrounded by intelligent people in their day-to-day life to be stimulated. I wouldn’t be surprised if the membership of a selective club for clever people is most sought after by people missing something from their lives.