Helpful as always, LinkedIn
I’ve frequently seen people on both the left and the right in the U.S. say “the other guys have successfully pushed their agenda further and further for decades, and we’ve been flailing about ineffectually over the same period.”
(I was reminded of the right-wing side of this the other day when I randomly decided to read “The Flight 93 Election,” by that pseudonymous “Journal of American Greatness” blogger whose schtick was being an “intellectual for Trump”)
Has anyone written about this phenomenon? I mean, there would be lots of boring things one could say about it, but there must also be some non-boring things that could be said – say about how much it is a difference in which facts are emphasized, vs. same facts getting interpreted differently, vs. the two sides not even agreeing about the facts.
[1801.00631] Deep Learning: A Critical Appraisal →
Gary Marcus, whose book The Algebraic Mind* I read ages ago, has a new paper critiquing deep learning. None of it is especially new or revelatory, but I do agree with a lot of it, and he cites some cool papers e.g. one that shows how the DeepMind Atari nets are brittle with respect to slight perturbations of game environments, in the course of proposing a new framework that is kinda similar to (although much more sophisticated than) the thing I came up with when I tried to attack the same problem myself
ETA: part about language does leave me wondering if he’s aware of Richard Socher’s work on recursive (not recurrent!) nets
ETA2: oh he cites Socher, but still says a moment later that all existing approaches have problems as shown by a paper that considered only recurrent nets. ???
*(more or less a critique of connectionism, i.e. the weird and not very impressive psychology modeling stuff neural net people were doing before deep learning became technically feasible and made them the new cool kids)
Also I’m still confused about why he refers to “experiments” in that binary digit generalization example (he did so in The Algebraic Mind as well) – one doesn’t have to do experiments to see that if you always set a certain input to 1 while training the net will never learn what that input does. Indeed, that is why it’s a useful example, because it’s so obvious what goes wrong!
[1801.00631] Deep Learning: A Critical Appraisal →
Gary Marcus, whose book The Algebraic Mind* I read ages ago, has a new paper critiquing deep learning. None of it is especially new or revelatory, but I do agree with a lot of it, and he cites some cool papers e.g. one that shows how the DeepMind Atari nets are brittle with respect to slight perturbations of game environments, in the course of proposing a new framework that is kinda similar to (although much more sophisticated than) the thing I came up with when I tried to attack the same problem myself
ETA: part about language does leave me wondering if he’s aware of Richard Socher’s work on recursive (not recurrent!) nets
ETA2: oh he cites Socher, but still says a moment later that all existing approaches have problems as shown by a paper that considered only recurrent nets. ???
*(more or less a critique of connectionism, i.e. the weird and not very impressive psychology modeling stuff neural net people were doing before deep learning became technically feasible and made them the new cool kids)
I think this is my favorite RLM video. “Emperor Palpatine Surgical Reconstruction Center” still gets me every time.
IIRC @ranma-official‘s (?) blog subtitle used to be “Eliezer Yudkowsky did not die for this,” and now that phrase pops into my head whenever any character talks about Apollo Mojave
I know I complained about this exact thing before, but I keep reading papers where researchers try to measure the duration of a phenomenon by finding the earliest time they can’t statistically detect it with p < .05
No! Nooooo! This is so bad in so many ways! It has multiple comparisons problems, it has within-vs-between-subjects problems (easily fixable by doing a paired test but they never do that), but those aren’t even the main problem, the main problem is that you’re making the duration a function of your sample size and it’ll get longer or shorter if you re-do the study with a different sample size
It’s striking to me how much most movies and TV – outside of the horror genre per se – shy away from physically disturbing content, even when they’re happy to take mentally disturbing content to extreme levels.
Like, I just saw a post making the argument “if you justify the frequency of rape in GoT with ‘that’s how it was in the middle ages,’ then why aren’t the characters getting dysentery all the time? why don’t they all have terrible teeth?” And I agree with this insofar as it’s an argument about something that was originally in text form, and agree that it’s a problem for the “historical accuracy” argument even on TV … but when you think about it, it’s obvious that “a non-comedic TV show where everyone has terrible teeth” is just culturally impossible, at least for the moment, in a way that has nothing to do with history (i.e. characters have equally good teeth in “historical” shows as in present-day ones, and for exactly the same reason).
Everyone in TV and movies is good-looking, except sometimes when they’re evil, but even then not usually. Maybe a non-evil character won’t be as good-looking if their ugliness has to be a plot point, but even then they’ll usually be “Hollywood ugly.” And when horrible things happen to people they’re usually skewed to be not that gross: people die of illnesses and gunshot wounds but we rarely see leaking bodily fluids, festering wounds, etc. except for some perfunctory blood to indicate “yes, they’ve been shot/stabbed.” Truly gross depictions of violence are limited to the ghetto of “torture porn,” which would make sense if violence itself was confined to torture porn, but no, it’s pervasive in regular movies, it just isn’t gross. And of course no one ever goes to the bathroom, unless there’s some plot point that has to happen there.
So people don’t have dysentery in popular TV shows, even popular TV shows where horrible calamities are commonplace, because horrible calamities are OK but gross stuff is not. Apparently. I’m honestly not sure why it is this way – obviously, gross stuff is unpleasant to look at, but TV and movies can offer plenty of other superficially unpleasant sensations (bleakness/sadness, awkwardness, fear, anxiety) in versions that people not only tolerate but even seek out, so IDK.
There is probably no name more liberally and more confusingly used in dynamical systems literature than that of Lyapunov (AKA Liapunov). Singular values / principal axes of strain tensor JJ (objects natural to the theory of deformations) and their longtime limits can indeed be traced back to the thesis of Lyapunov [ 10, 8], and justly deserve sobriquet ’Lyapunov’. Oseledec [8] refers to them as ‘Liapunov characteristic numbers’, and Eckmann and Ruelle [11] as ‘characteristic exponents’. The natural objects in dynamics are the linearized flow Jacobian matrix J t , and its eigenvalues and eigenvectors (stability exponents and covariant vectors). Why should they also be called ’Lyapunov’? The covariant vectors are misnamed in recent papers as ‘covariant Lyapunov vectors’ or ‘Lyapunov vectors’, even though they are not the eigenvectors that correspond to the Lyapunov exponents. That’s just confusing, for no good reason - Lyapunov has nothing to do with linear stability described by the Jacobian matrix J, as far as we understand his paper [10] is about JJ and the associated principal axes. To emphasize the distinction, the Jacobian matrix eigenvectors {e(j) } are in recent literature called ‘covariant’ or ‘covariant Lyapunov vectors’, or ‘stationary Lyapunov basis’ [ 12]. However, Trevisan [7] refers to covariant vectors as ‘Lyapunov vectors’, and Radons [ 13] calls them ‘Lyapunov modes’, motivated by thinking of these eigenvectors as a generalization of ‘normal modes’ of mechanical systems, whereas by ith ‘Lyapunov mode’ Takeuchi and Chat´e [ 14] mean {λj, e(j) }, the set of the ith stability exponent and the associated covariant vector. Kunihiro et al. [15] call the eigenvalues of stability matrix (4.3), evaluated at a given instant in time, the ‘local Lyapunov exponents’, and they refer to the set of stability exponents (4.7) for a finite time Jacobian matrix as the ‘intermediate Lyapunov exponent’, “averaged” over a finite time period. The list goes on: there is ‘Lyapunov equation’ of control theory, which is the linearization of the ‘Lyapunov function’, and the entirely unrelated ‘Lyapunov orbit’ of celestial mechanics.
(Cvitanović et. al., Chaos: Classical and Quantum)
I’ve never gotten far enough into this particular stuff to feel the brunt of this particular terminological clusterfuck, but I felt a stab of painful recognition reading this, because this sort of thing happens all the time
One of the little frustrations of grad school was simultaneously holding in mind (broadly justified) sense that all the profs and paper-authors were far, far smarter than me and the (also justified) sense that they were very sloppy writers and it was my job to clean up their thoughts inside my own head, because they weren’t going to do it for me
Nonlinear Dynamics 1: Geometry of Chaos →
what is this course about?
The theory developed here (that you will not find in any other course :) has much in common with (and complements) statistical mechanics and field theory courses; partition functions and transfer operators are applied to computation of observables and spectra of chaotic systems.
Nonlinear dynamics 1: Geometry of chaos (see syllabus)
- Topology of flows - how to enumerate orbits, Smale horseshoes
- Dynamics, quantitative - periodic orbits, local stability
- Role of symmetries in dynamics
Nonlinear dynamics 2: Chaos rules (see syllabus)
- Transfer operators - statistical distributions in dynamics
- Spectroscopy of chaotic systems
- Dynamical zeta functions
- Dynamical theory of turbulence
The course, which covers the same material and the same exercises as the Georgia Tech course PHYS 7224, is in part an advanced seminar in nonlinear dynamics, aimed at PhD students, postdoctoral fellows and advanced undergraduates in physics, mathematics, chemistry and engineering.
why this course
Most institutions have too few graduate students in any narrow speciality to offer a high level specialized course tailored to them. This Specialized Open Online Course is an experiment in sharing such advanced course with the off-campus research cohort. [A longer blurb]
prerequisites
A basic background in linear algebra, calculus, ordinary differential equations, probability theory, classical and statistical mechanics: ability to work with equations involving vectors and matrices, differentiate simple functions, and understand what a probability distribution is. Weekly homework assignments require both analytic and numerical work, so we will teach you Python as we go along. Knowledge of Matlab or Octave or another programming language is a very helpful. For introductory literature, check the book.