Koyon Ingantaccen Albarkatun Lissafi (CoRE-Learning): Tsarin Ka'idar Rarraba Lokaci don Koyon Injin

1. Gabatarwa & Dalili

Ka'idar koyon injin ta al'ada tana aiki a ƙarƙashin zato na ɓoye, wanda sau da yawa ba gaskiya ba: ana samun albarkatun lissafi marasa iyaka ko isassu don sarrafa duk bayanan da aka karɓa. Wannan zaton ya karye a cikin yanayin duniya na gaske kamar koyo daga rafi, inda bayanai ke zuwa ci gaba da yawa sosai. Takardar ta yi jayayya cewa aikin koyo ya dogara ba kawai akan yawan bayanan da aka karɓa ba, amma da mahimmanci akan yawan da za a iya sarrafa idan aka yi la'akari da ƙayyadaddun albarkatun lissafi—wani abu da ka'idar gargajiya ta yi watsi da shi.

Marubutan sun zana kwatance mai ƙarfi ga juyin halittar tsarin kwamfuta, suna bambanta wuraren "manyan kwamfutoci masu hankali" na yanzu (waɗanda ke ware ƙayyadaddun albarkatu na musamman ga kowane mai amfani/ aiki) da tsarin aiki na rarraba lokaci na zamani. Sun kawo waɗanda suka sami lambar yabo ta Turing Fernando J. Corbató da Edgar F. Codd don ayyana manufofin rarraba lokaci guda biyu: ingancin mai amfani (amsa cikin sauri) da ingancin kayan aiki (madaidaicin amfani da albarkatu ta hanyar tsarawa). Babban jigon shi ne cewa dole ne ka'idar koyon injin ta haɗa waɗannan damuwar rarraba lokaci, wanda ke kaiwa ga shawarar Koyon Ingantaccen Albarkatun Lissafi (CoRE-Learning).

2. Tsarin CoRE-Learning

Tsarin CoRE-Learning a hukumance yana gabatar da tsarawa da ƙuntatawar albarkatu cikin tsarin koyo. Ya watsar da garantin cewa za a iya sarrafa duk bayanai, yana mai da tsarin tsarawa ɗan ƙasa na farko a cikin ka'idar koyo.

2.1. Ra'ayoyin Tsakiya: Zaren Aiki & Nasara

Ana kiran aikin koyon injin da aka gabatar zuwa wata babbar cibiyar kwamfuta da zaren aiki. Kowane zaren aiki yana da ƙayyadaddun tsawon rayuwa tsakanin lokacin farawa da lokacin ƙarshe. Zaren aiki yana da nasara idan za a iya koyon samfuri wanda ya cika buƙatun aikin mai amfani a cikin wannan tsawon rayuwa. In ba haka ba, rashin nasara ne. Wannan tsarin yana haɗa sakamakon koyo kai tsaye zuwa ƙuntatawar lokaci da albarkatu.

2.2. Ƙarfin Aiki na Koyon Injin

An yi wahayi daga ra'ayoyin daga hanyoyin sadarwa da tsarin bayanai, takardar ta gabatar da ƙarfin aiki na koyon injin a matsayin ma'auni na zahiri don tsara tasirin albarkatun lissafi da tsarawa.

2.2.1. Ƙarfin Aiki na Bayanai

Ƙarfin aiki na bayanai ($\eta$) an ayyana shi azaman kashi na bayanan da aka karɓa da za a iya koya a kowane raka'a lokaci. Maɓalli ne mai ƙarfi wanda abubuwa biyu ke tasiri: yawan bayanan da ke shigowa da kasafin albarkatun lissafi da ake da su.

Takardar ta yarda cewa wahalar bayanai na iya bambanta (misali, saboda canjin ra'ayi, danganta da koyo a cikin buɗaɗɗen yanayi), tana ba da shawarar wannan a matsayin abu don haɗawa cikin samfurin ƙarfin aiki a nan gaba.

3. Tsari na Fasaha & Bincike

Duk da cewa ɓangaren PDF da aka bayar bai gabatar da cikakkun hujjojin lissafi ba, ya kafa tsarin da ake buƙata. Aikin algorithm na koyo $\mathcal{A}$ a ƙarƙashin CoRE-Learning ba kawai aiki ne na girman samfurin $m$ ba, amma na ingantaccen bayanan da aka sarrafa, wanda ke ƙarƙashin ƙarfin aiki $\eta(t)$ da manufar tsarawa $\pi$ akan lokaci $t$.

Wani sauƙaƙan tsari na hasarar da ake tsammani $R$ zai iya zama: $$R(\mathcal{A}, \pi) \leq \inf_{t \in [T_{\text{start}}, T_{\text{deadline}}]} \left[ \mathcal{C}(\eta_{\pi}(t) \cdot D(t)) + \Delta(\pi, t) \right]$$ inda $\mathcal{C}$ shine kalmar rikitarwa wanda ya dogara da adadin bayanan da aka sarrafa har zuwa lokacin $t$, $D(t)$ shine jimillar bayanan da aka karɓa, $\eta_{\pi}(t)$ shine ƙarfin aikin da aka samu a ƙarƙashin manufa $\pi$, kuma $\Delta$ kalmar hukunci ce don nauyin tsarawa ko jinkiri. Manufar ita ce a nemo manufar tsarawa $\pi^*$ wacce ta rage wannan iyaka a cikin tsawon rayuwar zaren aiki.

4. Tsarin Bincike & Misalin Lamari

Yanayi: Dandalin ML na gajimare yana karɓar zaren koyo guda biyu: Zaren A (rarrabuwar hoto) tare da ƙayyadaddun lokaci na awa 2, da Zaren B (ganowa abin da ba na al'ada ba akan rajistan ayyuka) tare da ƙayyadaddun lokaci na awa 1 amma tare da fifiko mafi girma.

Binciken CoRE-Learning:

1. Ayyana Zaren Aiki: Ayyana tsawon rayuwa, ƙimar zuwan bayanai, da manufar aiki ga kowane zaren aiki.
2. Samfurin Ƙarfin Aiki: Kimanta ƙarfin aikin bayanai $\eta$ ga kowane nau'in zaren aiki akan kayan aikin da ake da su (misali, GPUs).
3. Manufar Tsarawa ($\pi$): Kimanta manufofi.
   • Manufa 1 (Na Musamman/FCFS): Gudanar da Zaren A har zuwa ƙarshe, sannan B. Haɗari: Zaren B tabbas ya rasa ƙayyadaddun lokacinsa.
   • Manufa 2 (Rarraba Lokaci): Ware kashi 70% na albarkatu ga B na mintuna 50, sannan 100% ga A na sauran lokacin. Bincike ta amfani da samfurin ƙarfin aiki zai iya hasashen ko duka zaren aikin biyu za su iya cimma manufofin aikin su a cikin tsawon rayuwarsu.
4. Hasashen Nasara/Rashin Nasara: Tsarin yana ba da tushen ka'ida don hasashen cewa Manufa 1 tana haifar da rashin nasara ɗaya, yayin da Manufa 2 da aka tsara da kyau za ta iya haifar da nasara biyu, tana haɓaka ingancin kayan aiki gabaɗaya da gamsuwar mai amfani.

5. Ayyukan Gaba & Hanyoyin Bincike

• Horar da Babban Samfuri na Tushe: Tsara ayyukan horo na farko a cikin gungu masu bambanta (GPUs/TPUs) tare da farashin albarkatu mai ƙarfi (misali, AWS Spot Instances). CoRE-Learning na iya inganta ciniki tsakanin farashi da aiki.
• Haɗin Koyo na Gefe-Gajimare: Tsara sabuntawa na samfuri da ayyukan fassara tsakanin na'urorin gefe (ƙarancin wutar lantarki) da gajimare (babban ƙarfi) a ƙarƙashin ƙuntatawar bandwidth da jinkiri.
• MLOps & Koyo Ci gaba: Sarrafa tsarin sake horar da bututun a cikin tsarin samarwa lokacin da sabbin bayanai suka zo, tabbatar da sabuntawar samfuri ba tare da keta yarjejeniyar matakin sabis (SLAs) ba.
• Haɗawa da Koyo a cikin Buɗaɗɗen Yanayi: Faɗaɗa ra'ayin ƙarfin aiki $\eta$ don yin la'akari da ƙarfin aiki na wahala, inda farashin albarkatu kowane bayani yana canzawa tare da canjin ra'ayi ko sabon abu, haɗawa da fagage kamar ci gaba da koyo da gano abin da ba na al'ada ba.
• Iyakan Haɗuwa na Ka'ida: Samun garantin koyo irin na PAC waɗanda suka haɗa da kasafin kuɗi na albarkatu da manufofin tsarawa a sarari, ƙirƙirar wani sabon yanki na "ka'idar koyo mai ƙayyadaddun albarkatu."

6. Nassoshi

1. Codd, E. F. (Shekara). Take na aikin da aka ambata akan tsarawa. Mawallafi.
2. Corbató, F. J. (Shekara). Take na aikin da aka ambata akan rarraba lokaci. Mawallafi.
3. Kurose, J. F., & Ross, K. W. (2021). Sadarwar Kwamfuta: Hanya Ta Sama-ƙasa. Pearson. (Don ayyana ƙarfin aiki).
4. Zhou, Z. H. (2022). Koyon Injin a cikin Buɗaɗɗen Yanayi. Bita na Kimiyya ta Ƙasa. (Don haɗawa da canjin wahalar bayanai).
5. Silberschatz, A., Korth, H. F., & Sudarshan, S. (2019). Ra'ayoyin Tsarin Bayanai. McGraw-Hill. (Don ƙarfin aiki na ma'amala).
6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Cibiyoyin Adawa na Haɓakawa. Ci gaba a cikin Tsarin Bayanai na Jijiya. (Misalin tsarin ML mai cike da lissafi).
7. Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar Hotuna zuwa Hotuna marasa Haɗin gwiwa ta amfani da Cibiyoyin Adawa masu Daidaituwa na Zagaye. Taron Ƙasa da Ƙasa na Kwamfuta na Kwamfuta (ICCV). (Misalin aikin horo mai nauyin albarkatu).

7. Binciken Kwararru & Zargi