Ucwaningo lwe-Google Lethula i-SensorFM: Imodeli Yesisekelo Sempilo Esigqokekayo Esiqeqeshelwe Ngaphambilini Ngamaminithi Ayizigidi Eziyisigidi Eziyinye Yedatha Yezinzwa

Amamodeli amaningi ezempilo agqokekayo akhelwa umphumela owodwa ngesikhathi. Leyo ndlela ihlehla ezindaweni ezingamashumi amathathu nanhlanu. Amalebula ayabiza futhi isichasiselo se-retrospective singenakwenzeka.

I-Google Research yethuliwe I-SensorFM, imodeli eyisisekelo yezempilo egqokekayo eqeqeshwe kusengaphambili imizuzu engaphezu kwesigidigidi esingu-1 yedatha yenzwa evela kubantu abayizigidi ezingu-5.

Yini i-SensorFM?

I-SensorFM iyimodeli yesisekelo Senzwa Enkulu yokufunda ukumelela uchungechunge lwesikhathi olugqokekayo. Ingenisa izici ezihlanganisiwe ezingu-34 zomzuzu owodwa ezithathwe kuzinzwa ezinhlanu: i-PPG, i-accelerometer, i-EDA, izinga lokushisa lesikhumba, ne-altimeter. Lezo zici zihlelwe zaba izigaba eziyisikhombisa, efasiteleni lamahora angama-24.

Umgogodla uyisishumeki se-ViT-1D esiqeqeshwe ngomgomo we-masked-autoencoder kanye nosayizi wesichibi [20, 1]. Ukuqeqeshwa kwangaphambili kusebenzise ababambiqhaza abavunyelwe abangu-5,000,000, amasampula phakathi kuka-Septhemba 2024 no-Septhemba 2025. Leyo khophasi ihlanganisa amazwe angu-100+, zonke izifundazwe ezingu-50 zase-US, namamodeli angu-20+ e-Fitbit ne-Pixel Watch. Isamba samahora angaphezu kwezigidigidi ezimbili, noma imizuzu engaphezu kwesigidigidi esisodwa.

Kukhona okune okuhlukile, ngakunye kubhanqiwe ngevolumu yedatha elinganiselwe.

Ukuhlola kusebenzisa idatha ehlukene. Ihlanganisa izifundo eziyi-13,985 kuzo zonke izifundo ezintathu ezingase zigunyazwe yi-IRB. Lezo impilo ye-metabolic, yenhliziyo neyokuphefumula (N = 1,655), ukulala (N = 6,377), nempilo yengqondo (N = 5,953). Imisebenzi engama-35 ihlanganisa imithambo yenhliziyo (6), i-metabolic (8), impilo yengqondo (8), ukulala (3), izibalo zabantu (4), kanye nendlela yokuphila (6).

Icala Lokukala

Ngalokho kusetha, umbuzo wokuqala ngowokuthi isikali sithenga noma yini elinganisekayo. Ithimba labacwaningi lishanele amamodeli amasayizi amane liqhathaniswa namavolumu amane edatha.

I-SensorFM-B ku-5M corpus yehlisa ukulahlekelwa kokuqinisekisa ukwakhiwa kabusha ngo-31% uma iqhathaniswa ne-SensorFM-XXS. Ukulahlekelwa kokukhiqiza kwehla ngo-28% ngokwesilinganiso. Phansi komfula, izuza i-ΔAUC = 0.09 ekuhlukaniseni futhi Δr = 0.21 ekuhlehleni. Kuzo zonke izinhlobo ezihlukahlukene, u-B uwina imisebenzi engu-33 kwengu-35, futhi amazinga e-XXS agcina kokungu-33 kwangu-35.

Icala lokwehluleka linolwazi ngokufanayo. I-SensorFM-B eqeqeshwe ezifundweni ezingu-5K kuphela ithumela ukulahlekelwa kokuqinisekisa okungu-1.082. Lokho kubi kakhulu ukwedlula zonke izinhlobo ezincane zevolumu efanayo. Ukuqeqeshwa kwangaphambi kwesikhathi kuye kwamiswa ngaphambi kwesikhathi ngenxa yokuthi imodeli igcwele ngokweqile.

Ngakho-ke, yonke imiphumela yesihloko ithatha amavolumu edatha akalwe ngokulingana nomthamo. Ngokuhambisana naleyo diagonal enesilinganiso esilinganayo, kusho ukuthi i-ROC AUC ihamba .664, .681, .710, .752. Mean Pearson r unyakaza .386, .435, .536, .612. Isibalo esingenhla sibonisa ukuthi ithrendi ayikasuthi.

INHLOSO: Ukuphatha Idatha Elahlekile Njengesiginali

Ukukala kukodwa akuzichazi lezo zinombolo. Ukusakaza kwangempela kuyingxenye ngesikhathi sokushaja, izikhathi ezingasebenzi esihlakaleni, nezindlela zokonga amandla. Izindlela ezijwayelekile zifaka izikhala, zifake ukuchema, noma ziwise amawindi, zilahle idatha.

I-SensorFM esikhundleni salokho isebenzisa i-Adaptive and Herited Masking (AIM), eyethulwe ngu-Xu et al. ku-LSM-2. Imaski esetshenzisiwe iwukuhlanganisa imaskhi yokulahleka ezuzwe njengefa kanye nemaski yokwenziwa. Ukulahlekelwa kubalwa kuphela kumapeshi ambozwe ngokwenziwa abeneqiniso eliyisisekelo. Ukufihla amathokheni okunezigaba ezimbili, kusetshenziswa ithokheni eyehlayo kanye nokufihla ukunakwa, kugcina lokhu kusebenza kahle.

Ngoba idikhoda ifunda ukwakha kabusha okuphawulwe okwehlisiwe, ukuqiniswa nokubikezela kuza mahhala.

Ukwakhiwa kabusha kwe-MSE kusethi yokuhlola ebanjiwe, okuphansi kungcono.

Ngokumelene nesisekelo esingcono kakhulu, i-SensorFM ithuthukisa ukufaka okungahleliwe ngo-74.8%. Ukufakwa kwesignali yenzwa kuba ngcono ngo-83.7%.

Izandla Zivuliwe: Ukulungisa Okushumekiwe

Ukuguqula leso sifanekiso sibe izibikezelo kuqondile. Isifaki khodi sihlala sifriziwe. Ukushumeka kuhlanganiswe ngomuntu ngamunye, kusetshenziswa ukuchezuka kwencazelo nokujwayelekile kuzo zonke izinsuku. Lezo zinciphisa zibe izingxenye eziyinhloko ezingu-50. Ikhanda lomugqa libe seliqeqesha ngaphansi kokuqinisekiswa okuphindwe kahlanu, okuzimele komuntu.

import numpy as np
from sklearn.decomposition import PCA
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import roc_auc_score

def person_level(emb, pid):
    """Collapse day-level embeddings into one vector per participant."""
    people = np.unique(pid)
    feats = []
    for p in people:
        e = emb[pid == p]                    # (n_days, d)
        feats.append(np.concatenate([e.mean(axis=0), e.std(axis=0)]))
    return np.nan_to_num(np.stack(feats)), people   # pandas std() is NaN at 1 day

X, people = person_level(emb, pid)           # emb: frozen SensorFM embeddings
y = labels[people]                           # one label per participant

aucs = []
for tr, te in StratifiedKFold(5, shuffle=True, random_state=0).split(X, y):
    pca = PCA(n_components=50).fit(X[tr])    # PCA-50, fit on the train fold only
    clf = LogisticRegression(max_iter=400)   # paper: AdamW, lr 5e-3, wd 1e-4, 400 steps
    clf.fit(pca.transform(X[tr]), y[tr])
    p = clf.predict_proba(pca.transform(X[te]))[:, 1]
    aucs.append(roc_auc_score(y[te], p))
print(np.mean(aucs))

Lolu phenyo lomugqa lwehlula isisekelo esigadwayo sobunjiniyela besici kwimisebenzi engama-34 kwengama-35. Imiphumela ekhethiwe iyalandela.

Umugqa wokugcina awuwona ongaphandle. Izikolo ze-ASCVD ne-Framingham zibalwa kusukela kuzici zezibalo zabantu. Ngakho-ke amamodeli ezibalo zabantu kuphela awina ngokwakhiwa. Ithimba labacwaningi libika i-SensorFM ehamba phambili emisebenzini engama-31 kwengama-35, hhayi yonke.

Izixwayiso ezimbili zihlala ematafuleni afanayo. Izibalo zabantu zisasiza i-SensorFM emisebenzini engu-22 kwengu-30, nakuba ilifti incipha ngesikali. Ezimisweni ezinamalebula aphansi kakhulu, okubalulekile kwezibalo zabantu kukodwa kuhlala kuqinile.

Ikilasi le-Agentic

Ngisho nophenyo oluqondile ludinga ukulungiswa komsebenzi ngamunye. Ukwenza lokho ngokuzenzakalelayo, ithimba labacwaningi liqhube 'ikilasi' lama-ejenti abafundi abahlanu be-LLM. Lezi zi-span gemini-2.5 zikhanya ngokubuka kuqala kwe-gemini-3.1 pro. Ama-ejenti akhiqiza, akhiphe, athole amaphuzu, futhi acwengisise amakhanda ePython ngaphezu kwemijikelezo engu-20, esebenzisa ukushumeka okungancishisiwe.

Sekukonke benze izivivinyo eziyi-30,516. Amakhanda atholwe yi-agent ashaya i-probe yomugqa emisebenzini yokuhlukanisa engu-16 kwengu-20, elinganiswa ngo-F1. Baphinde bakhulisa ukuhlobana kwe-Pearson emisebenzini engu-12 kweyi-15 yokuhlehla. Ikhwalithi yesisombululo ilandelele i-Artificial Analysis Intelligence Index.

Izixazululo eziwinayo ziyalandela. Cishe zonke zinciphise isikhala sokushumeka saba yizilinganiso ezingama-50-100. Amamodeli alayini abemaningi kunawangenawo umugqa, futhi ama-ensemble avele ngaphansi kwekota.

Ukufaka i-ejenti yezempilo yomuntu siqu

Ukuhlolwa kokugcina kuhlola i-SensorFM njengethuluzi, hhayi ukufakwa kwebhentshimakhi. I-Gemini 3 Flash ekhiqizwe izifinyezo zezempilo zamaphrofayela ababambiqhaza bangempela angama-31. Yonke imibandela ithole izibalo zabantu kanye namamethrikhi ansuku zonke aklanyelwe isici. Izimo zibe sezingeza izibikezelo ze-SensorFM, okuqondiwe kweqiniso eliphansi, noma lutho.

Ngokusho kwephepha locwaningo, odokotela abane abanezitifiketi zebhodi, abaphuphuthekiswe isimo, bakhiqize izilinganiso ze-1,860 kuzo zonke izilinganiso zerubrikhi ezinhlanu. Ukungeza izibikezelo ze-SensorFM kwehlula inani eliyisisekelo lilonke (W = 10110, p <0.001), nakuhlangothi ngalunye. Izibikezelo zayo zazingenakuhlukaniswa ngokwezibalo kusukela eqinisweni eliyisisekelo (p = 0.396).

Sebenzisa Amacala

• Ukuhlola kanye nokuhlukaniswa kwengozi: Isifaki khodi esifriziwe kanye nekhanda elilodwa lomugqa lihlaba umkhosi kumakhandidethi omsebenzi wokuqinisekisa walebhu. Iphepha libheka lokhu ekuhlolweni, hhayi ukuxilongwa.
• Ukulungisa izifinyezo zansuku zonke: Njengoba kuhoxisiwe amaminithi angu-60 ahlanganayo, i-SensorFM igcina isibalo sezinyathelo esingu-99.7% kanye nokunemba okungu-99.9% kokulala okujulile.
• Izifundo ezishoda ngamalebula: Phenya ukushumekwa okufriziwe esikhundleni sokuqeqesha ekugcineni. Qhathanisa nesisekelo sezibalo zabantu kuphela kuqala.
• Ukuqeqeshwa okusekelwe phansi: Umyalo we-ejenti uyakwenqabela ukukhipha amanani angahluziwe okuhlehla noma amafulege we-boolean. Izibikezelo zihunyushwa ngokwekhwalithi kunalokho.

Isihloli Esisebenzisanayo