lsusr

• lsusr 25 May 2020 18:06 UTC

  3 points

  in reply to: MoritzG's comment on: Small Data

  NN also do not re­duce/​ideal­ize/​sim­plify, ex­plic­itly gen­er­al­ize and then run the re­sults as hy­poth­e­sis forks. Or use pri­ors to run checks (BS re­jec­tion /​ speci­fic­ity). We do.

  This is very im­por­tant. I plan to fol­low up with an­other post about the nec­es­sary role of hy­poth­e­sis am­plifi­ca­tion in AGI.

• lsusr 25 May 2020 18:03 UTC

  2 points

  in reply to: MoritzG's comment on: Small Data

  NNs are a big data ap­proach, tuned by gra­di­ent de­scent. Be­cause NNs are a big data ap­proach, ev­ery up­date is nec­es­sar­ily small (in the math­e­mat­i­cal sense of first-or­der ap­prox­i­ma­tions). When up­dates are small like this, av­er­ag­ing is fine. Espe­cially con­sid­er­ing how most neu­ral net­works use sig­moid ac­ti­va­tion func­tions.

  While this av­er­ag­ing ap­proach can’t solve small data prob­lems, it is perfectly suit­able to to­day’s NN ap­pli­ca­tions where things tend to be well-con­tained, with­out fat tails. This ap­proach works fine within the tra­di­tional prob­lem do­main of neu­ral net­works.

• lsusr 21 May 2020 21:39 UTC

  2 points

  in reply to: Mark Xu's comment on: What are ob­jects that have made your life bet­ter?

  Does the pen re­quire cus­tom ink car­tridges or does it ac­cept stan­dard ones?

• lsusr 21 May 2020 21:38 UTC

  2 points

  on: What are ob­jects that have made your life bet­ter?

  Un­in­stal­lable product place­ment sparks hate when­ever I see it. Cus­tom ROMs for my phone (LineageOS and Re­s­ur­rec­tion Remix) dra­mat­i­cally im­prove my life over com­mer­cial phone op­er­at­ing sys­tems be­cause cus­tom ROMs avoid con­tam­i­nat­ing my at­ten­tion with things I don’t want.

  1. There are no un­in­stal­lable ap­pli­ca­tions serv­ing as ad­ve­tise­ments.

  2. They do not try to sell my at­ten­tion such as via Google News.

  In ad­di­tion, it is easy to get root ac­cess to these phones. That makes it easy to in­te­grate it with my Linux lap­top.

Orthogonality

• lsusr 20 May 2020 22:27 UTC

  7 points

  on: Is there a way to repli­cate pho­to­syn­the­sis ar­tifi­cially?

  Chloro­phyll en­ters an ex­cited state when the right fre­quency of light in­ter­acts with it. Pho­to­voltaics op­er­ate on a similar prin­ci­ple, though it would be a stretch to say they were de­signed “in the image” of chloro­phyll.

  Pho­to­voltaics har­ness sun­light more cheaply and effi­ciently than man­u­fac­tured chloro­phyll. Once the sun­light is har­nessed, you can do what­ever you want with it such as se­ques­ter­ing car­bon diox­ide.

  The only ad­van­tage chloro­phyll-based sun­light har­vest­ing has over pho­to­voltaics is that plants are easy to grow. Man­u­fac­tur­ing chloro­phyll in a lab­o­ra­tory would buy the in­effi­ciency of plants at the price of in­or­ganic man­u­fac­tur­ing. It’s the worst of both wor­lds.

  On the other hand, chloro­phyll is used as a method of car­bon se­ques­tra­tion. This is called “grow­ing forests”.

• lsusr 16 May 2020 6:12 UTC

  −5 points

  in reply to: Grotace's comment on: Crav­ing, suffer­ing, and pre­dic­tive pro­cess­ing (three char­ac­ter­is­tics se­ries)

  Once we have deemed that want­ing to pur­sue plea­sure and hap­piness are wire­head­ing-like im­pulses, why stop our­selves from say­ing that want­ing to im­pact the world is a wire­head­ing-like im­pulse?

  Is there a way you can rephrase this ques­tion with­out us­ing the word “wire­head”? When dis­cussing med­i­ta­tion, the word “wire­head” can have two very differ­ent mean­ings. Usu­ally, “wire­head” refers to the gross failure mode of heavy med­i­ta­tion where a prac­ti­tioner amnes­the­sizes him/​her­self into a potato. Kaj_So­tala has used the word “wire­head” to re­fer to a sub­tle spe­cific con­se­quence of taṇhā.

  Why isn’t a de­sire to avoid death crav­ing?

  A de­sire to avoid death is crav­ing. (In fact, death is one of the Four Sights.) The ac­tions of post­pon­ing death are not crav­ing. Only the de­sire to avoid death is.

  You clearly speak as if go­ing to a den­tist when you have a tooth ache is the right thing to do, but why?

  Be­cause you have a toothache and your teeth will rot if you don’t go to a den­tist.

  Once you dis­tance your ‘self’ from pain, why not dis­tance your­self from your rot­ting teeth?

  Pen­e­trat­ing taṇhā is the op­po­site of dis­tanc­ing. It’s about ac­cept­ing the world right now as it is. If your teeth are rot­ting right this in­stant then you should ac­cept that your teeth are rot­ting right this in­stant. Such is the Li­tany of Tarski.

  If my teeth are rot­ting,

  then I de­sire to be­lieve my teeth are rot­ting;

  If my teeth are not rot­ting,

  then I de­sire to be­lieve my teeth are not rot­ting;

  Let me not be­come at­tached to be­liefs I may not want.

  The thing you dis­tance your­self from isn’t the pain, it’s your self. Kaj_So­tala’s post is about taṇhā, one of the Three Char­ac­ter­is­tics of Ex­is­tence. Another Char­ac­ter­is­tic of Ex­is­tence is anattā or non-self. I hope this be­comes clearer once Kaj_So­tala gets to anattā in this se­ries.

  All my in­tu­itions about how to act are based on this flawed sense of self. And from what you are out­lin­ing, I don’t see how any in­tu­ition about the right way to act can pos­si­bly re­main once we lose this flawed sense of self.

  It is pos­si­ble to do some­thing with­out crav­ing it. For ex­am­ple, con­sider re­lax­ing on a trop­i­cal beach and reach­ing over to drink a mango smoothie. Now, con­sider the in­stant you are mid-sip, suck­ing through the straw while the fla­vor washes over your mouth. In that in­stant, you act with­out crav­ing.

  The same goes for when you are en­grossed in fun con­ver­sa­tion with close friends and fam­ily.

  There’s a gen­eral dis­com­fort I have with this se­ries of posts that I’m not able to fully ar­tic­u­late, but the above ques­tions seem re­lated.

  Good!

• lsusr 15 May 2020 4:04 UTC

  2 points

  in reply to: eapache's comment on: Small Data

  Yes.

• lsusr 14 May 2020 19:10 UTC

  4 points

  in reply to: crabman's comment on: Small Data

  I think in mod­ern ma­chine learn­ing peo­ple are ex­per­i­ment­ing with var­i­ous in­duc­tive bi­ases and var­i­ous ad-hoc fixes or tech­niques with help cor­rect­ing for all kinds of bi­ases.

  The con­clu­sion of my post is that these fixes and tech­niques are ad-hoc be­cause they are writ­ten by the pro­gram­mer, not by the ML sys­tem it­self. In other words, the cre­ation of ad-hoc fixes and tech­niques is not au­to­mated.

• lsusr 14 May 2020 18:48 UTC

  2 points

  in reply to: crabman's comment on: Small Data

  In your ex­am­ple with a non-con­verg­ing se­quence, I think you have a typo—there should be $MSD\left(t\right)$ rather than $MSD\left(x_t\right)$

  Fixed. Thank you for the cor­rec­tion.

• lsusr 14 May 2020 17:16 UTC

  2 points

  in reply to: maximkazhenkov's comment on: Ma­chine Learn­ing Can’t Han­dle Long-Term Time-Series Data

  Which one of their songs has a re­peated cho­rus? I could not iden­tify one in the Elvis Presley rock song nor the Kay Perry pop song.

Small Data

• lsusr 14 May 2020 4:19 UTC

  2 points

  in reply to: maximkazhenkov's comment on: A game de­signed to beat AI?

  In ad­di­tion to your ideas, I would add long tales and chaotic sys­tems. It’s hard to train an AI on 1,000,000 dat­a­points when the value func­tion of the 1,000,001st dat­a­point could out­weigh all the pre­vi­ous cu­mu­la­tive re­sults.

  To gen­er­al­ize this math­e­mat­i­cally, a board game ought to have a value func­tion that never con­verges no mat­ter how many times you play the game.

• lsusr 14 May 2020 3:47 UTC

  2 points

  on: A game de­signed to beat AI?

  Ma­chine learn­ing is bad at situ­a­tions where it is pro­vided with limited train­ing data. There­fore I would de­sign a game with fre­quently-chang­ing rules. In par­tic­u­lar, I would cre­ate an ex­pan­sion set for Be­trayal at House on the Hill.

  Be­trayal at House on the Hill is about ex­plor­ing a haunted house. The gim­mick is you do not know the rules of the game be­fore you be­gin play­ing. There are many differ­ent rules the haunted house might obey.

  Hu­mans could beat ma­chines for a long time if the fol­low­ing two er­atta were ap­plied:

  1. An in­tel­li­gence may not have ac­cess to in­for­ma­tion ahead of time about the ex­pan­sion packs’ rules. (It is fair play for the AI’s pro­gram­mers to have ac­cess to the base rule­set but not the spe­cial iter­a­tion-spe­cific rule­sets.)

  2. An in­tel­li­gence only gets points for win­ning the first time it en­coun­ters a par­tic­u­lar rule­set.

  An AI would need to read the spe­cial­ized rules on-the-spot and then un­der­stand the se­man­tics well enough to de­vise a strat­egy. Then the com­puter would have to ex­e­cute this strat­egy cor­rectly on its first try. No soft­ware in ex­is­tence to­day can do any­thing like this.

  Not only could hu­mans crush ma­chines at this board game, to­day’s best ma­chine learn­ing soft­ware can­not even play this game (fol­low the rules) with­out its pro­gram­mers’ read­ing the com­plete rule­book ahead of time, which is cheat­ing.

  As for goal #2, Be­trayal at House on the Hill is my fa­vorite board game.

• lsusr 14 May 2020 3:30 UTC

  2 points

  in reply to: seed's comment on: A game de­signed to beat AI?

  You could achieve a similarly lop­sided re­sult faster with the in-per­son Tur­ing test, albeit at the ex­pense of goal #2.

  Rules

  Each player guesses whether the op­po­nent is hu­man or ma­chine. Each player gets one point for a cor­rect guess and one point for con­vinc­ing the op­po­nent that he/​she/​it is a hu­man.

  To beat hu­mans at this, we would need the fol­low­ing de­vel­op­ments:

  1. A robotic body that can perform the same func­tions as hu­man biol­ogy.

  2. A robotic body aes­thet­i­cally in­dis­t­in­guish­able from a hu­man body.

  3. An ar­tifi­cial mind that can con­vinc­ingly simu­late hu­man be­hav­ior.

Re: Some Heroes

[Question] Plans for COVID-19?

• lsusr 1 May 2020 19:21 UTC

  1 point

  on: SlateS­tarCodex 2020 Pre­dic­tions: Buy, Sell, Hold

  Good luck! Not gonna jinx it.

  That’s nice of you.

• lsusr 29 Apr 2020 18:44 UTC

  5 points

  in reply to: NaiveTortoise's comment on: I do not like math

  To me, math­e­mat­ics is like lift­ing weights. I don’t like lift­ing weights. I do it any­way, half­heart­edly and with poor form, be­cause I need to move fur­ni­ture some­times.

  I am ex­trin­si­cally mo­ti­vated to do math. So why didn’t I ma­jor in ap­plied math­e­mat­ics? I wor­ried “ap­plied math­e­mat­ics” would be wa­tered down. Be­sides, I want my pro­fes­sors to be rigor­ous. Read­ing and writ­ing proofs of known facts is a higher level of rigor than mat­ters to me most of the time. On the other hand, writ­ing proofs for new the­o­rems is lots of fun.

  Math is a logic of words founded on ab­solute truth. I be­lieve in nei­ther words nor ab­solute truth. I can play by the rules of logic, but deep down I think prob­a­bil­is­ti­cally.

I do not like math