Overview
Gumbel-Softmax idhiri rinoita kuti neural network 'sample' kubva kune discrete mapoka ichiri kudzidziswa ne gradient descent. Izvo zvine basa nekuti backpropagation kazhinji haigone kuyerera kuburikidza nekusarudzika, discrete sarudzo.
Gumbel-Softmax uye Reparameterization inyanzvi yekuvaka inobata mhando yemhando, mutengo wezvivakwa, latency, uye kuvimbika pachiyero.
Deep Dive
Neural network inodzidza nekutumira gradients kumashure kuburikidza nekushanda kwese. Asi sampling chikamu chakasarudzika (sekunhonga izwi # 7 pa50,000) yakaoma, isingasiyanise kusvetuka, saka magradients anofira ipapo. Iro reparameterization trick inonyora zvekare sampling isina kujairika kuitira kuti kusarongeka kuuye kubva kune yakamisikidzwa yekunze ruzha sosi, ichisiya yakatsetseka, inosiyanisa nzira yemagradients. Gumbel-Softmax inoshandisa izvi kumhando dzakasiyana-siyana: inowedzera Gumbel-yakaparadzirwa ruzha kune zvinyorwa, yobva yatsiva iyo yakaoma argmax ine tembiricha-inodzorwa softmax. Pakupisa kwepamusoro kubudiswa kunoputika kwakatsetseka pamusoro pezvikamu; sezvo tembiricha inodonha yakananga ku zero inorodza yakananga kune imwe-inopisa vector, kudzoreredza sampling chaiyo uchiramba uchisiyaniswa kwese.
Technical Insight
Iyo Gumbel-Max trick inoti: kuwedzera yakazvimiririra Gumbel(0,1) ruzha kune yega yega logit uye kutora iyo argmax inoburitsa chaiyo sampuli kubva mukugovera softmax. Gumbel-Softmax inochinjanisa iyo yakaoma argmax ye softmax((log p + g)/tau). Tembiricha tau inopindirana pakati pekugadzika, kwepamusoro-entropy kugovera (tau hombe) nepedyo-discrete imwe-inopisa (diki tau). Nekuti iyo ruzha g inotorwa kunze kwetiweki, nzira kubva pamalogi kuenda kune inobuda inoramba ichisiyaniswa.
Mastering Gumbel-Softmax uye Reparameterization
Gumbel-Softmax idhiri rinoita kuti neural network 'sample' kubva kune discrete mapoka ichiri kudzidziswa ne gradient descent. Izvo zvine basa nekuti backpropagation kazhinji haigone kuyerera kuburikidza nekusarudzika, discrete sarudzo. Gumbel-Softmax uye Reparameterization inyanzvi yekuvaka inobata mhando yemhando, mutengo wezvivakwa, latency, uye kuvimbika pachiyero. Kuti uvake kunzwisisa kwakadzama, bata Gumbel-Softmax uye Reparameterization semuenzaniso wekushandisa, kwete chinhu chimwe chete: tsanangura zvaunoda mhedzisiro, kujekesa fungidziro, uye patsanura izvo zvinogona kuitwa nehurongwa hwakavimbika kubva kune zvichiri kuda kutonga kwenyanzvi.
Mukuita, zvikwata zvakasimba zvinoshandisa Gumbel-Softmax uye Reparameterization inokwidziridza zvivakwa, data, uye sarudzo dzezvivakwa zvinopesana nekuvimbika uye mutengo. Ivo vanonyora zvakajeka maitiro ebudiriro, bvunzo vachipokana ne data rechokwadi uye mafambiro ebasa, uye iterate zvichibva pane zvakacherechedzwa maitiro ekutadza kwete kuhwina-nguva imwe chete yebhenji. Apa ndipo apo kunzwisisa kwe theoretical kunoshanduka kuve kugona kwakasimba pane chigadzirwa, mutemo, uye mashandiro.
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore. Panguva imwecheteyo, Kukwirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba. Nzira yakatsiga ndeyekubatanidza kukurumidza kuyedza nekutonga: mhanyisa vatyairi vendege, tora humbowo, buritsa matanda esarudzo, uye urambe uchivandudza chengetedzo semaitiro emuenzaniso, zvinotarisirwa nemushandisi, uye zvinodikanwa zvekutonga.
Strategic Impact
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore.
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete.
Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira.
Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Real-World Implementation
Kudzidzira akasiyana-siyana autoencoder ane categorical (discrete) latent macode pane anongoenderera eGaussian.
Differentiable neural architecture search (semuenzaniso, DARTS-style nzira) kusarudza kuti ndeipi mashandiro ekuisa pane imwe neimwe layer.
Kudzidza discrete codebook sarudzo muVQ-maitiro uye discrete anomiririra modhi.
Yakasiyana-siyana nzira kana gating sarudzo mumusanganiswa-we-nyanzvi uye mamiriro-computation network.
Maitiro Ekuita
Gumbel-Softmax uye Reparameterization mukuita
Kudzidzira akasiyana-siyana autoencoder ane categorical (discrete) latent macode pane anongoenderera eGaussian.
Kudzidzira akasiyana-siyana autoencoder ane categorical (discrete) latent kodhi panzvimbo yekungoenderera chete maGaussian mamwe Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura mhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa nemitengo yekukanganisa nekufamba kwenguva.
Gumbel-Softmax uye Reparameterization mukuita
Differentiable neural architecture search (semuenzaniso, DARTS-style nzira) kusarudza kuti ndeipi mashandiro ekuisa pane imwe neimwe layer.
Differentiable neural architecture yekutsvaga (semuenzaniso, DARTS-maitiro maitiro) kusarudza kuti ndeipi mashandiro ekuisa pane yega yega Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando zvikumbaridzo kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Gumbel-Softmax uye Reparameterization mukuita
Kudzidza discrete codebook sarudzo muVQ-maitiro uye discrete anomiririra modhi.
Kudzidza discrete codebook sarudzo muVQ-maitiro uye discrete anomiririra modhi Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Gumbel-Softmax uye Reparameterization mukuita
Yakasiyana-siyana nzira kana gating sarudzo mumusanganiswa-we-nyanzvi uye mamiriro-computation network.
Yakasiyana nzira kana sarudzo dzekugezera mumusanganiswa-we-nyanzvi uye ane mamiriro-makomputa network Zvikwata zvinowanzowana mhedzisiro iri nani pazvinenge zvichitsanangudza mhando kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Njodzi & Guardrails
Kugadzirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba.
Infrastructure uye mari yekugadzirisa inowanzotarisirwa pasi.
Chengetedzo uye kucherechedzwa mapundu anogona kukura sezvo masisitimu anowedzera kuoma.
Implementation Roadmap
Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa.
Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Benchmark pasi pechokwadi mutoro uye data mamiriro.
Benchmark pasi pechokwadi mutoro uye data mamiriro. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro.
Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera.
Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.