Google p\u0159ich\u00e1z\u00ed s modelem TabFM (Tabular Foundation Model), kter\u00fd slibuje z\u00e1sadn\u00ed zjednodu\u0161en\u00ed pr\u00e1ce s tabulkov\u00fdmi daty. A\u017e dote\u010f platilo, \u017ee pokud jste cht\u011bli z firemn\u00edch datab\u00e1z\u00ed (nap\u0159. pro predikci odchodu z\u00e1kazn\u00edk\u016f nebo detekci podvod\u016f) dostat spolehliv\u00e9 v\u00fdsledky, museli jste s\u00e1hnout po algoritmech typu XGBoost. To znamenalo hodiny a dny manu\u00e1ln\u00edho lad\u011bn\u00ed parametr\u016f (hyperparameter tuning) a p\u0159\u00edpravy datov\u00fdch rys\u016f (feature engineering). TabFM naproti tomu vyu\u017e\u00edv\u00e1 princip „in-context learning“, zn\u00e1m\u00fd z velk\u00fdch jazykov\u00fdch model\u016f. Tr\u00e9noval se na stovk\u00e1ch milion\u016f syntetick\u00fdch tabulek a dok\u00e1\u017ee analyzovat slo\u017eit\u00e9 vztahy v datech takzvan\u011b „zero-shot“ \u2013 tedy okam\u017eit\u011b, v jedin\u00e9m kroku a bez nutnosti model jakkoliv dotr\u00e9nov\u00e1vat pro konkr\u00e9tn\u00ed dataset. Google jej nav\u00edc brzy integruje p\u0159\u00edmo do sv\u00e9ho datov\u00e9ho skladu BigQuery, kde k pokro\u010dil\u00e9 predikci posta\u010d\u00ed jednoduch\u00fd SQL p\u0159\u00edkaz. <\/p>\n\n\n\n
Koment\u00e1\u0159:<\/p>\n\n\n\n
N\u00e1pad aplikovat logiku, kterou zn\u00e1me z ChatGPT, na nudn\u00e9 excelov\u00e9 \u010di datab\u00e1zov\u00e9 tabulky zn\u00ed z pohledu b\u011b\u017en\u00e9ho byznysu naprosto snov\u011b. Odstran\u011bn\u00ed „\u010dern\u00e9 magie“ spojen\u00e9 s lad\u011bn\u00edm parametr\u016f by mohlo otev\u0159\u00edt pokro\u010dilou datovou analytiku i men\u0161\u00edm firm\u00e1m, kter\u00e9 si nemohou dovolit platit drah\u00e9 t\u00fdmy data scientist\u016f. M\u00e1 to ale jeden z\u00e1sadn\u00ed h\u00e1\u010dek \u2013 TabFM je tr\u00e9nov\u00e1n \u010dist\u011b na syntetick\u00fdch, um\u011ble vygenerovan\u00fdch datech, proto\u017ee sd\u00edlet re\u00e1ln\u00e1 pr\u016fmyslov\u00e1 data je kv\u016fli citlivosti t\u00e9m\u011b\u0159 nemo\u017en\u00e9. Ot\u00e1zkou tedy z\u016fst\u00e1v\u00e1, jak si model porad\u00ed s „\u0161pinav\u00fdmi“ daty re\u00e1ln\u00e9ho sv\u011bta \u2013 s chyb\u011bj\u00edc\u00edmi hodnotami, lidsk\u00fdmi chybami p\u0159i zad\u00e1v\u00e1n\u00ed nebo s nestandardn\u00edmi form\u00e1ty, se kter\u00fdmi si zku\u0161en\u00ed analytici u tradi\u010dn\u00edch model\u016f musej\u00ed ru\u010dn\u011b poradit. Funkce „zadej SQL p\u0159\u00edkaz a dostane\u0161 v\u00fdsledek bez n\u00e1mahy“ zn\u00ed skv\u011ble pro management, ale re\u00e1ln\u011b hroz\u00ed, \u017ee u\u017eivatel\u00e9 bez hlub\u0161\u00edch znalost\u00ed strojov\u00e9ho u\u010den\u00ed za\u010dnou d\u011blat byznysov\u00e1 rozhodnut\u00ed na z\u00e1klad\u011b predikc\u00ed, u kter\u00fdch v\u016fbec nebudou tu\u0161it, jak k nim model dosp\u011bl. <\/p>\n\n\n\n
Zdroj:<\/p>\n\n\n\n