tel

47 ms·

271.
by tel 7y ago
JAX is similar to many other autodifferentiation frameworks being developed for machine learning. It's a spiritual successor to the popular autograd framework (same authors). The advantages JAX brings are * Numpy-adjacent interfa
272.
by tel 7y ago
For many kinks, it doesn't in practice seem to matter that much. Most applications are computing large stochastic derivatives which smooth out kinks through averaging. Critical kinks are those that affect the geometry of the gradient i
273.
by tel 7y ago
Many frameworks already compute derivatives, but they don't use a symbolic representation. Instead they use a method called "automatic differentiation" which does something along the lines of (a) extracts a trace of the algor
274.
by tel 7y ago
This is a really important post, if you're the kind of person who thinks about this sort of thing. Iteration speed isn't the endpoint, just a measurement. Learning speed is more important, but only if you're also using th
275.
by tel 7y ago
Prototypical inheritance (which, I assume, is the other side of what you're referring to) is also weird. Prototypes eliminate classes as a special thing (relegating them to mere constructors or factories, which is fair) but they don&#x
276.
by tel 7y ago
If classes were just that then they'd not be so bad. They also tend to bring inheritance along with them, though, and that's a messy beast.
277.
by tel 7y ago
This sort of formalism is bad for handling multiple returns and thus exceptions. For instance, they become very simple in continuation passing style whereas GOTO is difficult. Flowcharts feel like a simple and universal way of talking about
278.
by tel 7y ago
Heck, I still think of variables as boxes with labels.
279.
by tel 7y ago
I strongly like the idea of pioneers as compared to "settlers" and "town planners" (see Simon Wardley's writing). But the idea that pioneers have and/or benefit from no process is BS. The best pioneer is struct
280.
by tel 7y ago
You’re getting all kinds of answers but a key one, for me, is that film is unavoidably slow and expensive. Digital cameras teach you that each frame is cheap and throwaway. That encourages some kinds of photographic/artful thinking whi
281.
by tel 7y ago
I think this is key: there was a massive learning effort that went into the current dplyr interface. This risked fragmenting the community, but Hadley and his collaborators navigated those waters effectively. That's super tough work an
282.
by tel 7y ago
I really, really dislike all the dtype wrangling and how those choices resonate throughout the API. I understand that a lot of work has been done to make that API "work", but in practice it feels like that effort would have been b
283.
by tel 7y ago
This falls apart when computation is a factor, though. You can't recompute the whole notebook on every commit when there are 30 cells that each take 2-8 seconds to complete.
284.
by tel 7y ago
My summary: Neural nets have massive "capacity" which means that, in the face of finite data sets, they can both (a) reasonably represent generalizable and non-generalizable functions and (b) can take on priors which do not distin
285.
by tel 7y ago
That's not entirely true. The point being made is about the consequences of the design being set up with that notation. That design and that notation is reasonably general. It requires some familiarity with notation around mathematical
286.
by tel 7y ago
Sound and unsound aren’t the same as strong and weak. Weak and strong systems aren’t honestly very well defined, but under many definitions can be both sound and unsound, though the theory of the weak system will need to have explicit inclu
287.
by tel 7y ago
You handle continuity in discrete spaces through topology. A function is continuous for a (really, two) given topology if the inverse images of open sets are open sets.
288.
by tel 7y ago
The punchline at the end is worth calling out: Turing Equivalence applies at first-order computations like `Nat -> Nat`, but fails at higher-order computation like `(Nat -> Nat) -> {0, 1}`.
289.
by tel 7y ago
Weekly billing doesn’t mean you work 40 hours, though. Additionally, one can imagine taking a week off or prorating a month.
290.
by tel 7y ago
Why do you suggest hourly? I understand fixed big puts all of the risk on you, but would you also consider daily or weekly billing as a reasonable compromise?
291.
by tel 7y ago
Honestly, it's not surprising. ML is billed as a tool, one that in the last 5 or so years we've surprisingly "figured out". This is vast overselling, but it still creates the basic mental model for folks without further
292.
by tel 7y ago
In my opinion, there's space for a "ML Product Manager" as a specialization for someone who understands the technical aspects of both software and ML systems, but also can design roadmaps, build stakeholder buy-in, and genera
293.
by tel 7y ago
Ha, that's good to hear. Would love to chat with you about it if you're interested.
294.
by tel 7y ago
There are a load of questions here. > Where do I go from here? Keep trying different models? > ...after [the labeling is] all done is this even going to work? > [How to label] > If we spend all this time and money labelling foot
295.
by tel 7y ago
Speculation only, but the method of proof might, in its key step, only bound to ~200. This is a more meaningful result to present if your reader knows that 200 is a as good as 1 because it loses nothing and also highlights something about t
296.
by tel 7y ago
Conda can manage a few things beyond just Pip. Things like binaries, system libs, R installs. I'm not entirely convinced that's a good thing, but it's an easy path to managing those sorts of libraries alongside your Python li
297.
by tel 7y ago
I think this is nice. It's a method of calculation derived from an important fact—that quadratic equations have two solutions and those solutions can be written in such a form. I don't have the experience or interest to debate whe
298.
by tel 7y ago
First up, what you seem to be talking about is Deep Learning, not Machine Learning in general. In more general ML there are many theorems, some also apply to DL. Also, the step of "shuffle around the ML graph using some intuition"
299.
by tel 7y ago
If computation is the process of spending effort to receive information, then memory is partially paid-down computation. There are lots of ways to blur these lines.
300.
by tel 7y ago
Isn't this derived from conservation of momentum? As I understand it, that arises from a symmetry in spatial configuration: that physics behaves identically along shifts in position. I'm well beyond anything I'm familiar with

More ›