Incredible Machine Learning
-
Not credible, to be more precise:
Machine learning is booming in medicine. It's also facing a credibility crisis
Researchers analyzed hundreds of papers on machine learning models developed to combat the Covid-19 pandemic. They found every single one was fatally flawed.
The use of statistical models for stuff like this is kind of interesting but ultimately not surprising that you can't just throw lots of stuff at it and get magic out the other end.
-
@boomzilla said in Incredible Machine Learning:
Not credible, to be more precise:
Machine learning is booming in medicine. It's also facing a credibility crisis
Researchers analyzed hundreds of papers on machine learning models developed to combat the Covid-19 pandemic. They found every single one was fatally flawed.
The use of statistical models for stuff like this is kind of interesting but ultimately not surprising that you can't just throw lots of stuff at it and get magic out the other end.
Oh, you get magic alright...
-
@boomzilla as much as I hate machine learning, this is in principle a perfect use case: looking at complex data and then coming up with lots of hypotheses about which features in the data are correlated with the illness that a human wouldn’t even begin to consider. Of course then you need to actually verify these hypotheses to see if green jelly beans are really are significant.
In ML the first step of this is usually a split of test and training data. But you need a large, representative data set to begin with and it appears that’s one point where all of these studies fail already.The replication crisis is even more problematic in ML than in other fields, where everyone just puts out their magic ML models and nobody has the data or code to even consider if that’s possibly correct:
The only possible recourse is IMO what many researches have suggested: not only have open access papers but also provide a repository with both code and data to produce the exact output that was claimed.
But as mentioned in the article, that’s problematic with medical data.
-
@topspin said in Incredible Machine Learning:
@boomzilla as much as I hate machine learning, this is in principle a perfect use case: looking at complex data and then coming up with lots of hypotheses about which features in the data are correlated with the illness that a human wouldn’t even begin to consider
Does that ever really happen with ML? Seems like it remains a black box. How would it help (in this case) trained radiologists develop and test hypotheses that simply having them review stuff wouldn't?
-
@boomzilla said in Incredible Machine Learning:
@topspin said in Incredible Machine Learning:
@boomzilla as much as I hate machine learning, this is in principle a perfect use case: looking at complex data and then coming up with lots of hypotheses about which features in the data are correlated with the illness that a human wouldn’t even begin to consider
Does that ever really happen with ML? Seems like it remains a black box. How would it help (in this case) trained radiologists develop and test hypotheses that simply having them review stuff wouldn't?
It might find things they have been trained to ignore. I expect you not to understand this.
-
@boomzilla there is research going on into “interpretable machine learning”. How well that works I’m not sure, maybe what I wrote was too optimistic in that regard.
But even if it stays a black box, you should find ways to ensure that the feature your algorithm has identified is actually relevant, by whatever method. Adding more data and more verification, I guess.
My point was I’m sure there can be features in the data that are both significant for diagnosing the illness and picked up by ML but not humans. But we need some way to make sure the models actually work once we think they might work, and aren’t just picking up noise or confounding variables. Otherwise you end up with:
“AI trained to classify skin lesions as potentially cancerous learns that lesions photographed next to a ruler are more likely to be malignant.”
-
@Gribnit said in Incredible Machine Learning:
@boomzilla said in Incredible Machine Learning:
@topspin said in Incredible Machine Learning:
@boomzilla as much as I hate machine learning, this is in principle a perfect use case: looking at complex data and then coming up with lots of hypotheses about which features in the data are correlated with the illness that a human wouldn’t even begin to consider
Does that ever really happen with ML? Seems like it remains a black box. How would it help (in this case) trained radiologists develop and test hypotheses that simply having them review stuff wouldn't?
It might find things they have been trained to ignore. I expect you not to understand this.
I understand that you misunderstood my question.
-
@topspin said in Incredible Machine Learning:
My point was I’m sure there can be features in the data that are both significant for diagnosing the illness and picked up by ML but not humans. But we need some way to make sure the models actually work once we think they might work, and aren’t just picking up noise or confounding variables.
Yeah. I'm just saying that the interpretable part seems really badly needed. And maybe you simply can't reliably do it when you've got some kind of neural net statistical aggregator thing going on. But maybe someone will figure it out.
FTR, I got that link from this:
https://blogs.sciencemag.org/pipeline/archives/2021/06/02/machine-learning-deserves-better-than-this
And he mentioned some of that stuff, which could be included in @Gribnit's statement (he'll deny it, of course):
That point is addressed in this recent preprint, which shows how such radiology analysis systems are vulnerable to this kind of short-cutting. That’s a problem for machine learning in general, of course: if your data include some actually-useless-but-highly-correlated factor for the system to build a model around, it will do so cheerfully. Why wouldn’t it? Our own brains pull stunts like that if we don’t keep a close eye on them. That paper shows that ML methods too often pick up on markings around the edges of the actual CT and X-ray images if the control set came from one source or type of machine and the disease set came from another, just to pick one example.
-
@boomzilla said in Incredible Machine Learning:
The use of statistical models for stuff like this is kind of interesting but ultimately not surprising that you can't just throw lots of stuff at it and get magic out the other end.
The problems with ML, as I see it, are of three kinds:
The first one is that they are, as you say, statistical models. That means they (may) give a good answer on average (or with a high probability), but not necessarily in all cases. That's not an issue when failure (either false positive or false negative) has a relatively low individual cost (e.g. serving the wrong ad to a given browsing profile), but becomes a huge issue when some individual outcomes are catastrophic (e.g. if the ML stuff says "patient is healthy" and no one looks any further).
For example, in my field of work, ML is getting more and more used to highlight anomalous data points that are then manually investigated. Erroneously flagging a data point as anomalous when it's not is not a big deal (the person looking at it just loses a bit of time), missing one point isn't really a big deal either (for decades we've developed various ways to deal with noisy data, so if a bit of noise remains in the data it won't break everything). This is a huge improvement compared to previous techniques where a few generic plots were used to try and identify outliers (but sometimes a point isn't an outlier on the plots that you have time to look at), or where a human was looking at some random sub-samples one-by-one to try and identify the type of errors that happened in that data set (and obviously spent a lot of time looking at good data, and could miss the bad one).
OTOH, despite a huge lot of research papers trying to do so, we don't really use ML to compute the outcome of complex processes, because this is where we want more than just a "statistical best guess." Those processes rely on modelling more or less complex physical phenomenon and sometimes the point that behaves differently from the rest (i.e. that doesn't follow the statistical law) is actually the needle that we're trying to find in the haystack of data!
The second issue is the "black box" aspect, and the fact that it's hard to know why a ML algorithm picked a specific solution. That one is fairly well-known and some work is being done on that, so maybe that'll improve with time.
The third one was initially overlooked by a lot of CS specialists who thought they could replace any domain-specific expert, and that's the fact that you still need an expert to both train, and interpret, the ML stuff. I remember an article some time ago where some nasty CFD was solved with some ML approach, but the breakthrough happened when they had the idea to feed frequency-transformed data -- that step required a domain expert to know that frequencies made sense for the specific problem they were tackling.
A similar example from my field is one case where we tried to find which spatial attribute(s) would be the best to predict another (think e.g. trying to predict the temperature in a room by feeding it a 3D grid of all physical things you can measure, from air flow, distance to features such as window or radiators, but also colour of the floor or whatever). The code told us that the best predictor was the x-coordinate, but what it entirely missed was that there was some strong geometric feature aligned with X (think e.g. an opened window at x=0 and a radiator at x=10). So in a sense it produced the "right" answer, but you still needed an expert to look at the result and say "no, there is no way that X is actually the driving variable, there is another hidden variable that we have to find to truly understand what's happening."
On the whole, I think ML is sliding into the through of the usual U-shaped curve of adoption. We're (thankfully) getting past the hype-peak, and interest is somewhat falling. Your article is a good illustration of many people falling off that peak. Hopefully what will now emerge are the few cases where it truly makes sense.
-
@remi said in Incredible Machine Learning:
We're (thankfully) getting past the hype-peak, and interest is somewhat falling. Your article is a good illustration of many people falling off that peak. Hopefully what will now emerge are the few cases where it truly makes sense.
Here's to hoping that's the case.
I'm still seeing a lot of attempts at "look we're MLing this slightly different thing" (frequently regardless of whether or not it makes sense). Getting off-peak will be good; right now there's too much volume to follow ML stuff sensibly, so whatever little good work may exist is swamped by the noise.
-
@boomzilla said in Incredible Machine Learning:
Does that ever really happen with ML? Seems like it remains a black box. How would it help (in this case) trained radiologists develop and test hypotheses that simply having them review stuff wouldn't?
The problem is that the input data had in-band salience markers that were being picked up on, rather than whatever is in the real data.
In general though, much of ML is just a fancy form of curve fitting. (Except it's not in 2D and it's a hypersurface instead.) As any good statistician will tell you, when doing curve fitting you want the order of the curve you fit to be as low as possible as that does the best things when presented with novel data. Otherwise you can go to almost arbitrary lengths to get an excellent fit for your sample datapoints, yet get totally bogus results for other values. You don't want to match the noise in the input data! It is the need to minimise the complexity order of the resulting models that most naïve ML users forget, and that's the cause of almost all the nasty input sensitivity problems that ML is infamous for. If you can keep the amount of over-fitting low, ML is extremely good at pattern matching and prediction.
Real brains have similar problems too, except that deep sleep has the effect of losing those high order bits; I suspect that's one of the most important functions of sleep. (This has been observed in small scale simulations using Hodgkin-Huxley neuron models for a small group of neurons.)
-
@remi said in Incredible Machine Learning:
For example, in my field of work, ML is getting more and more used to highlight anomalous data points that are then manually investigated. Erroneously flagging a data point as anomalous when it's not is not a big deal (the person looking at it just loses a bit of time), missing one point isn't really a big deal either (for decades we've developed various ways to deal with noisy data, so if a bit of noise remains in the data it won't break everything). This is a huge improvement compared to previous techniques where a few generic plots were used to try and identify outliers (but sometimes a point isn't an outlier on the plots that you have time to look at), or where a human was looking at some random sub-samples one-by-one to try and identify the type of errors that happened in that data set (and obviously spent a lot of time looking at good data, and could miss the bad one).
Yeah, my brother in law works on this kind of stuff for JPL, where they use it to have, say, the orbiter look through pictures it's taken and then decide what looks potentially interesting for humans to look at. Obviously, bandwidth is precious so they can't just send everything, but have had pretty good results with this approach.
-
@boomzilla said in Incredible Machine Learning:
@Gribnit said in Incredible Machine Learning:
@boomzilla said in Incredible Machine Learning:
@topspin said in Incredible Machine Learning:
@boomzilla as much as I hate machine learning, this is in principle a perfect use case: looking at complex data and then coming up with lots of hypotheses about which features in the data are correlated with the illness that a human wouldn’t even begin to consider
Does that ever really happen with ML? Seems like it remains a black box. How would it help (in this case) trained radiologists develop and test hypotheses that simply having them review stuff wouldn't?
It might find things they have been trained to ignore. I expect you not to understand this.
I understand that you misunderstood my question.
I expect nothing less.
-
@dkf said in Incredible Machine Learning:
it's the ordinary differential equation (ODE) from Hell
From what I remember of my Diff Eq class (practically nothing), that describes all of them.
-
@dkf statisticians are finding that the evils of overfitting are overstated.
-
From OP article:
“It was a real eye-opener and quite surprising how many methodological flaws there have been,” said Ian Selby, a radiologist and member of the research team.
Those eyes must've been closed really hard that they didn't notice it years ago.
-
@Gąska said in Incredible Machine Learning:
From OP article:
“It was a real eye-opener and quite surprising how many methodological flaws there have been,” said Ian Selby, a radiologist and member of the research team.
Those eyes must've been closed really hard that they didn't notice it years ago.
You can be well aware of the sorry state of many papers and still be surprised it’s worse than you thought. There’s no upper limit to being cynical.
-
@Captain said in Incredible Machine Learning:
Whoever downvoted me needs to learn about the bias-variance trade off.
-
@HardwareGeek said in Incredible Machine Learning:
@dkf said in Incredible Machine Learning:
it's the ordinary differential equation (ODE) from Hell
From what I remember of my Diff Eq class (practically nothing), that describes all of them.
Used to be I could bore my son away by starting to describe an integral. Calculus no longer scares the lad so I've switched to DiffEq as a droning topic to bore him away with. Odds are not in my favor, I do not know the subject.
-
@Captain said in Incredible Machine Learning:
@Captain said in Incredible Machine Learning:
Whoever downvoted me needs to learn about the bias-variance trade off.
It sounds like you have an interesting point to make. Could you elaborate instead of speaking in riddles?
-
There are philosophical and practical reasons to not necessarily accept the "axioms" of classical statistics of continuous random variables, especially when you're dealing with huge collections of categorical data, which is "essentially" noiseless. There isn't "any" measurement error to be had, since a categorical value ends up in one of N categories.
This is unlike the classical case, where there is data from a "continuous" domain, and there is the assumption of normally distributed measurement errors (ha!).
If you overfit in a classical case (say, fitting a line to data), you end up with the wrong line and the further away from the data points you get, the worse the model gets.
If you overfit in a broadly categorical case, you end up mostly making absolutely correct predictions, given the data, and sometimes end up making incorrect predictions where the training set merely didn't have enough data to fully represent a category (or a "case", as it were).
Basically, the machine uses probability to map out a domain's logic, and learns to make logical inferences.
(You can somewhat compare this to how kids learn grammar. They say stuff like "He swimmed across" because they "overfit" on one specific rule. But that doesn't represent a problem. It represents mastery of that rule. It's 'swam' that is the exception, and a new rule that needs to be learned. Which of course requires more data that actually includes the exception, so that the poor kid can learn it.)
I don't want to get too technical into the bias-variance trade off, but the basic idea is that you can make estimators with lower variance if you accept higher bias. Yes, that means higher mean error -- in the case above, if you get an inference wrong, you'll end up in the wrong category. But it's better than the alternative where the every estimate is a linear combination of mutually exclusive alternatives, and doesn't actually make a guess.
The B-V trade off gets used a lot, like in ridge regression and LASSO (1996! -- these aren't new ideas), two algorithms for biased linear least squares estimation with many parameters. They "recursively" add the most significant parameters to the model until the model is happy.
All that said, the "linear combination of mutually exclusive alternatives" definitely has its place and it gets used a lot, especially in neural networks and a bunch of linear least squares algorithms on categorical data.
-
@Captain Someone is butthurt. :-)
-
@Captain it’s clear from your post that you know what you’re talking about, but it still reads to me like you’ve got overfitting and underfitting swapped. Or maybe use it in a completely different way than I’m used to.
To take a simple example, let’s say that you’ve got something like Y=aX^2 + bX + error, i.e. a quadratic function with noisy data. From what I’ve understand of your post, you’d say that if I try to fit this data with a linear function I’m overfitting the data (or I’m just misunderstanding your post, very well possible). For me that’s underfitting, while overfitting would be if you learn a model that’s has many more than the two parameters a and b so that you don’t learn just the actual function but also bake in all the noise you found. That gives great performance on the training data, because you can basically memorize everything you’ve seen, but doesn’t generalize.
And that can easily happen if you take a NN with thousands of parameters and let it learn from 70 or so images of Covid patients. It’s like giving it to a kid who instead of learning the ask decides to memorize the patient names, then aces the test.
Of course there’s methods to prevent that. (Which you obviously know, I’m just mentioning it)
-
@topspin said in Incredible Machine Learning:
To take a simple example, let’s say that you’ve got something like Y=aX^2 + bX + error, i.e. a quadratic function with noisy data.
Overfitting would be if you learn a model that’s has many more than the two parameters a and b so that you don’t learn just the actual function but also bake in all the noise you found.
I think that would be over parameterization, which also would lead to overfitting.
I think what @Captain is talking about is even if you've got data reasonably well described by a quadratic model, you're often better off with a quadratic model that fits the data less well than it's possible to do so, if your goal is prediction of future, unseen observations. Basically, take the parameters that produce the lowest error on the data, and back off of them a bit.
Of course, exactly how you arrive at these biased "underfittted" parameters matters, you can't underfit the data any old which way and expect good generalization to new cases. For linear models, that's where stuff like lasso and ridge regression (previously mentioned) come into play.
-
@HannibalRex said in Incredible Machine Learning:
I think what @Captain is talking about is even if you've got data reasonably well described by a quadratic model, you're often better off with a quadratic model that fits the data less well than it's possible to do so, if your goal is prediction of future, unseen observations. Basically, take the parameters that produce the lowest error on the data, and back off of them a bit.
The core problem with ML is that it's often used with datasets where we don't have a good model of the ground truth or a strong handle on how much noise is present. (We may well have a model of the technical noise in the sensors, but not in the model-level noise of parameters that make basically no difference.) Because of that, and the way ML works, the primary danger (provided we don't simply not have enough data in the first place) is matching the noise. However, there's one useful trick: the noise usually has higher frequency than the model (i.e., more variation between samples). By relaxing the model — reducing the amplitude of higher-order frequencies in the fourier series, or whatever other technique you choose — you tend to get a better model. (It's equivalent to trying to pick the correct k in k-means clustering; higher values will give better fits, but there's a critical value below which fits are much worse, and that can be seen as a change in the angle of the graph of how well things fit, an essentially entropic measure.)
You should do other things too. For example, splitting your input data randomly into a training set and an evaluation set (usually in a ratio of 2:1) so that you can evaluate whether the model works on things it hasn't learnt from. This is a minor variation on technique used in quite a few sciences.
My usual criticism of standard machine learning is that it requires a crazy quantity of computation to do the learning rule, requiring very large input datasets and large numbers of repetitions. That feels horribly wrong, even though the results are often really quite good. There's got to be some trick that people are missing (I suspect it is in the connectivity matrix level).
-
@dkf said in Incredible Machine Learning:
I suspect it is in the connectivity matrix level
Have you tried talking to Morpheus?
-
@HardwareGeek It's common to describe the connections between simulated neurons as a 2D array (the connection matrix) that says what the strength of connection between each neuron is. Logically, a large value says that the two neurons are correlated with one causing the other to increase in activity, a zero says they're not actually connected, and a negative value says that one suppresses activity in the other. (Some models don't use suppression, biologically realistic models definitely do use it; suppression is a real thing and is critical for selectivity and homeostasis.) A lot of ML models start out by assuming that the connectivity matrix is basically flat, with everything connected to everything and the weights tuned so that “energy” isn't lost between layers of neurons, and then are evolved from there. But that's a very bogus state. The evidence from biology is that neurons alter their connectivity, their synapses, in response to how active they have been, and this connectivity change is a far more binary thing than many assume (there's synaptic strength too, but it's a less important factor it seems) and while I don't know the details of how much this has been published, neuromorphic systems that change their connectivity matrices substantially seem to be better at learning things.
The wholesale changing of the connectivity matrix is something that it is intensely difficult to do in silicon hardware. Reconfiguring networking hardware while it is live is where things get scary.
-
@dkf Your neurons seem to be deficient at detecting (intended) humor.
-
@HardwareGeek said in Incredible Machine Learning:
Your neurons seem to be deficient at detecting (intended) humor.
I blame the martini and the wine.
-
@dkf said in Incredible Machine Learning:
@HardwareGeek said in Incredible Machine Learning:
Your neurons seem to be deficient at detecting (intended) humor.
I blame the martini and the wine.
Just tell him you were out pendanting the pendant.
-
-
This could go in the spam emails thread but the particular "Fuck, No!" nature of this makes me think it also deserves a place here.
-
@topspin Considering how much language and structure varies in different fields, I'm fairly confident in saying that this will be as shit as something like Grammarly (wish I never had to see one of their ads again). Also, what's up with the name?
-
@cvi said in Incredible Machine Learning:
Also, what's up with the name?
What do you mean?
Enago sounds like electronic nagging, which is apt, Trinka sounds like "you'd have to be drunk to think this is a good idea."
Seriously, if I want to crank out garbage papers badly enough to use AI for it, I'd go for the old trusted SCIgen.
-
Frankly, at first glance that's not a hugely catastrophically bad idea. You sometimes see papers that are so badly written that anything would help to make them slightly more readable.
OTOH, it's likely that the tool will learn, and perpetuate, some of the worst writing quirks (*), and will only make things worse in the end.
(*) like "never use active form" (or "never use 'we'") which leads to awful passive-form-everything. Not only this makes sentences much more awkward to read, but it allows for weasel-wording everything. No, "the parameter" did not "had its value chosen" by some abstract and unspecified entity. You "chose the parameter", man-up to it and admit it. If that makes you cringe because you can't justify that choice, well maybe it's a good thing you notice it before the paper is actually published?
-
@remi said in Incredible Machine Learning:
No, "the parameter" did not "had its value chosen" by some abstract and unspecified entity. You "chose the parameter", man-up to it and admit it. If that makes you cringe because you can't justify that choice, well maybe it's a good thing you notice it before the paper is actually published?
You presume too much. The parameter might have been chosen by some undergrad that does not deserve to be credited.
Also, abstract and unspecified entity has a special name: The Holy Ghost.
-
@Kamil-Podlesak said in Incredible Machine Learning:
@remi said in Incredible Machine Learning:
No, "the parameter" did not "had its value chosen" by some abstract and unspecified entity. You "chose the parameter", man-up to it and admit it. If that makes you cringe because you can't justify that choice, well maybe it's a good thing you notice it before the paper is actually published?
You presume too much. The parameter might have been chosen by some undergrad that does not deserve to be credited.
Also, abstract and unspecified entity has a special name: The Holy Ghost.
They are assumed to have chosen in the case of any variable related to theology, theogony, theomachy, or string theory. All other variables are chosen by The Invisible Hand of the Market.
-
@remi said in Incredible Machine Learning:
Not only does this make
ssentences much more awkward to read,Since we're discussing grammar corrections...
-
@HardwareGeek said in Incredible Machine Learning:
@remi said in Incredible Machine Learning:
Not only does this make
ssentences muchmoreawkwarder to read,Since we're discussing grammar corrections...
FTFY
-
@remi said in Incredible Machine Learning:
(*) like "never use active form" (or "never use 'we'") which leads to awful passive-form-everything. Not only this makes sentences much more awkward to read, but it allows for weasel-wording everything. No, "the parameter" did not "had its value chosen" by some abstract and unspecified entity. You "chose the parameter", man-up to it and admit it. If that makes you cringe because you can't justify that choice, well maybe it's a good thing you notice it before the paper is actually published?
There should always be a minimum of at least one paragraph in active voice in a paper (excluding papers that are explicitly just summarizing the field) being the part where you claim that you did the work. Without that, it could just as well be done by someone else and then why the fuck are you writing a paper about it you plagiarizing bastard?!
-
@dkf Yes,
I think this rule is stupidthe author thinks this rule is stupidit is thought that this rule is stupid.And even a summary paper can use the active voice, for example when the authors explain how they categorised previous work ("we classify previous works as Method A or Method B because we consider this and that characteristic to be the most important distinction..."). A summary can never be perfectly objective and neutral so it's better to embrace it clearly. IMO there are very few cases where using the passive voice doesn't make things worse, either for readability or accountability, or both.
-
@remi Turgid scientific writing is one of my big dislikes about how science is communicated. And it's cargo cult/traditions all the way down--people do it now because that's what people have done. A lot of the early 20th century papers, even for complex subjects, were easy to read and had personality showing through. That got lost somewhere.
-
@Benjamin-Hall scientists are frequently not good writers (or if they are, it's more by coincidence than design), so they're all copying the badly-written style they see from all of the other brilliant-scientists-but-terrible-writers. I have written a couple of papers, and I perpetuated the bad writing a bit -- not egregiously bad, just run-of-the-mill scientific-writing-bad. Because that's the style, and I didn't even have a Master's degree yet, so I did what everyone else did.
Now that I'm more of an adult, I'd probably inject a little more human into a scientific paper, were I to write one. But not too much, you know -- can't let the humans know that the robots can act like them.
-
@remi said in Incredible Machine Learning:
*) like "never use active form" (or "never use 'we'") which leads to awful passive-form-everything.
Funnily enough, we have the opposite problem. It's almost to the point where if you have a single sentence of passive anywhere, there's a good chance somebody will complain about that. Yeah, I get it, 100% passive everywhere isn't nice to read and leads to some of the strange things mentioned. But forbidding all passive just because some people misunderstood a stupid style guide from about 100 years ago isn't quite the right thing either.
Anyway, instead of weird passive where parameters set themselves, there ends up being a lot of confusing use of "we" (the authors), "we" (the authors and the reader, who are apparently on this journey together) and "we" (fuck knows who, but at least it isn't in passive voice).
One of the better papers I tend to remember was written by a single author. They used "I" instead of "we". That was nice.
-
@Benjamin-Hall said in Incredible Machine Learning:
A lot of the early 20th century papers, even for complex subjects, were easy to read and had personality showing through. That got lost somewhere.
There's a tendency to over-complicate things, especially when the actual work is kinda mediocre. One example I see way too often is papers slapping down some (almost nonsensical) equations and definitions for seemingly no reason. This might be a bit of a computer science problem; it almost seems like some people have this inferiority complex that requires them to show off that they can use greek letters in their writing.
It's not very impressive if you're actually not scared off by big equations.
-
@PotatoEngineer said in Incredible Machine Learning:
Now that I'm more of an adult, I'd probably inject a little more human into a scientific paper, were I to write one.
I occasionally end up writing them. Not my favourite activity, and I'd much rather leave that to people who are officially academics (it's part of their job, after all) but sometimes you've got to muck in and get things moving or an important part of the story simply doesn't get told. When I do, I tend to make sure at least one paragraph in each section is active (unless it is truly describing background work).
Now, if only I could persuade more of my colleagues that doing lots of citations was a good thing, that'd be great. (I think it's a really nice gambit to ask a conference for an extra page because of the number of citations. Makes the editors of the proceedings really uncomfortable to say no…)
-
@dkf said in Incredible Machine Learning:
I think it's a really nice gambit to ask a conference for an extra page because of the number of citations. Makes the editors of the proceedings really uncomfortable to say no…
Chances are you'd get an extra page if you asked for it, especially if it's for something like references. Extra page is pretty easy these days. I've been told that I can approve that as an IPC member/primary reviewer without asking upstairs. Things aren't typically printed, so page counts aren't as important any more. Goal is to keep people from writing 20 page papers when 6 pages would do -- writing concisely is more difficult than verbose, so some like to skip the extra thought that is needed to optimize the text.
Not counting references in the first place is starting to happen here and there.
That said, don't just copy-paste those BibTeX entries (or whatever). Go over the damn references, and make sure they look alright and are consistent (abbreviating the venues the same way etc).
-
@dkf said in Incredible Machine Learning:
Now, if only I could persuade more of my colleagues that doing lots of citations was a good thing, that'd be great.
My bachelor's thesis had like 20 citations, only 5 or 6 of which were actually relevant. But I guess the incentives are different when you aren't writing a paper just to graduate.
As for passive voice, it was strictly enforced that we use it everywhere. It was especially annoying considering that I was describing what I've put into my app.
-
@cvi said in Incredible Machine Learning:
That said, don't just copy-paste those BibTeX entries (or whatever). Go over the damn references, and make sure they look alright and are consistent (abbreviating the venues the same way etc).
Sorta off topic, but when I did my PhD dissertation and sent it in for college-level editorial approval, I got rejected the first time. Why?
Because in two out of N (I forget how many) references in the reference section, I used (wait for it) two spaces after a period instead of one.
That was the only rejection reason. <sarc>Money well spent, I tell you what.</sarc>
-
@Benjamin-Hall said in Incredible Machine Learning:
@cvi said in Incredible Machine Learning:
That said, don't just copy-paste those BibTeX entries (or whatever). Go over the damn references, and make sure they look alright and are consistent (abbreviating the venues the same way etc).
Sorta off topic, but when I did my PhD dissertation and sent it in for college-level editorial approval, I got rejected the first time. Why?
Because in two out of N (I forget how many) references in the reference section, I used (wait for it) two spaces after a period instead of one.
That was the only rejection reason. <sarc>Money well spent, I tell you what.</sarc>
Now you get PRs rejected because you used tabs instead of spaces. See, it was good training!
