I actually wonder if somebody used a fake identity to set up an account with a warehousing/shipment fulfillment company that stocks things and ships them, then set up the appropriate EDI pipeline to send shipping orders to it... What would be the results if a decently budgeted adversary made something attractive looking that shipped malicious USB flash drives to anyone that requested one.
I know we're not in the era when a windows pc will happily run any autorun.inf and .EXE file found on an inserted flash drive or DVD anymore. But even so. What if it didn't even have any malicious data payload but somebody was shipping USB-A interface capacitor based usb killers?
What if it did have data on it and came with a slick color brochure walking people through how to run the binary, or in a linux or developer specific audience, how to 'sudo' the ELF binary that lives on its filesystem?
Even worse than a rubber ducky: the O.MG cable does the same thing, but looks like a regular USB cable. Their Apple Lightning-dupe [0] is my favorite. The creator was on Darknet Diaries a while ago, too. [1]
That's almost 12 years now. A novice can now get ATmega32 USB devices Prime delivered. Not a cutting edge theoretical attack anymore but a basic tool in a every pen tester's toolbox now.
The while premise of "learn some stuff them take an exam on exactly that stuff" is pretty flawed, and that's the point. So much of the academic structure is about what's convenient for evaluation, rather than what's best for learning. Why not get rid of the exam and replace it with something else entirely? Who says we have to have exams at all?
> So much of the academic structure is about what's convenient for evaluation, rather than what's best for learning
Academia is going to have to focus on the latter anyway because we have AI, which is exactly the same: it’s only useful when you can evaluate the output for accuracy independently. But how do you do this without throwing out objectivity entirely? Maybe we need to move to apprenticeships?
The problem is simplification. Suppose two regions share a border with some nonlinear points a, b, c, d. Simplifying the polygon for the first region might yield a, b, d while the second yield a, c, d. This creates gaps or overlaps between the two regions.
But what is the border? Set the border to what it actually is, not a simplification of it. The state of Colorado is formally a 697 sided polygon, don't simplify it to a rectangle.
This is not what OP is describing. It is very common to simplify objects for decreasing boundary objects by orders of magnitude. GeoJSON is missing correlation when you do that. Simplifying country objects from a GeoJSON source could lead to a gap between the country borders. So you either have poor representation or a longer pipeline to convert objects to an amenable object set. It also breaks idempotency in some regards.
To do the simplification, you detect shared borders, simplify and generate polygons again.
That doesn’t make topojson inherently superior. You can convert back and forth and for many applications geojson is easier to process.
Yes, you could write code to do that. Or use the utilities provided in the TopoJSON GitHub and let them do it for you: convert to TopoJSON, simplify, convert back to GeoJSON. They have already written all the code for you.
It depends on what purpose you are using the polygons. In an online map you need to simplify way down. Consider these Colorado maps at two different zoom levels:
Even the one zoomed in on the state appears to use maybe 15-20 vertices max.
In the second one, if I squint real hard I can just barely make out one slight dogleg on the western border and one on the south. And that is partly because I knew to look for them in the zoomed-in map.
If we use, say, the Census TIGER/Line boundary definitions for the states, we are probably talking about hundreds of thousands of vertices, perhaps millions. You won't be using those in an online map without simplifying.
The Texas border with Mexico is formally down the centerline of the Rio Grande, even as the river moves (ignoring fiddly complications). Even if you could somehow take a perfect snapshot of it at a given time, you'd run into the coastline paradox when sampling it.
So don’t simplify the shapes on their own. Geojson is a storage and exchange format, you can still convert it to other formats if you want to modify it.
I think what the original comment is pointing out is that GeoJSON lacks a concept of a shared boundary. Shared boundaries expressed in GeoJSON get around that by duplicating data. Whenever data is duplicated, there's a risk that the copies will not be exactly the same. That makes the task of modification more challenging given that the real world is full of messy data, like duplicates not matching.
20-25 years ago I worked a lot with map data from otherwise high quality, and sometimes authoritative, sources like the USGS and NOAA that had this non-identical shared boundaries problem (in formats other than GeoJSON). If the format doesn't allow such mistakes to be expressed, then they have to fix their data to publish it in said format.
Sure, but not every format is useful for everything. Geojson is great if you want a simple way to express a shape to show on a map. It’s like criticizing CSV because people put strings in choice value fields instead of doing a foreign key to another table. That’s just not what the format is used for.
I'd take your point further... No format is useful for everything. But we have to be aware of the trade-offs of each format (or language or tool or ...) in order to make the right choice of what to use for a given use case. We do that by sharing knowledge of where a given tool succeeds and where it falls down. Pointing out something a format doesn't handle well is not condemning that format for all use cases (I happily choose GeoJSON over other formats for many things).
I dunno... gradient descent is only really reliable with a big bag of tricks. Knowing good initializations is a starting point, but recurrent connections and batch/layer normalization go a very long way towards making it reliable.
I agree, this is the correct way to see it IMO. Instead of designing better optimizers, we designed easier parameterizations to optimize. The surprising part is that these parameterizations exist in the first place.
Gradient descent is mathematically the most efficient optimization strategy (safe for some special functions) in high dimensions. This goes so far that people nowadays even believe it has to be used in the human brain [1], if only because every other method of updating the brain would be way too energy inefficient. From that perspective, finding the right parameterization was all we ever needed to achieve AI.
Even in supervised ML, pure gradient descent is not the most efficient optimization strategy. E.g., momentum is ubiquitous, and the updates it induces cannot be expressed as a gradient of some scalar loss. But the rotational non-gradient component of its updates substantially improves performance and convergence on the architectures we use.
The brain probably primarily uses something like TD for task learning, which is also not expressible as a gradient of any objective function. And, though the paper mentions Hebbian learning, it's only very particular network architectures (e.g. single neuron; symmetric connections) that you can treat its updates as a gradient of some energy function; these architectures aren't anything close to what we see in the brain.
Pure gradient descent is not what happens in either field, but e.g. momentum is just another parameter constructed from historic gradients. While it is unlikely that the brain runs backpropagation the way you see it implemented in modern ML (same goes for TD btw), the core principle kind of needs to be the same from a pure large scale, high dimensional network efficiency POV. On top of that, adaptive plasticity is almost by definition about estimating useful directions of change. The key insight here would be that the brain does gradient estimation quite cheap and we can probably still learn a thing or two about modern ML from it.
Their claim isn't that the brain uses gradient descent, but that the direction of updates has (on average) positive inner product with the gradient. I expect this would also be true for (say) simulated annealing, yet we don't say that simulated annealing is gradient descent.
There's also a discussion of loss functions and how they relate to the update missing - as far as I know, there's still no great notion of how the brain picks a global loss function, and no mechanism for backprop. In this paper, looking at a specific learning task you can define a loss function extrinsically allowing us to talk about the gradient, but how that relates to things happening in the brain is a big big mystery.
The gradient is the direction in which loss improves the fastest. Moving in a direction with a positive dot product with the gradient just means that you're (locally) improving the loss.
Hmm I'm not sure what you mean by "Gradient descent is mathematically the most efficient optimization strategy". Do you mean that gradient-based optimization in general? (in other words do you consider Adam gradient descent?)
Probably 'we reserve the right to train our next version of smart autocomplete based on the text you send to the current version of smart autocomplete'
Which is not different in kind from “we use your source code to improve our products” and is functionally identical to “we own your output because you use our editor.”
How do people continually fall for this. Refusing to look at the playbook that has been run time and time again and then getting offended when it is too late.
reply