52: Machine Learning at the UTC-Heudiasyc lab

Pro­fes­sor Philippe Bon­ni­fait, also Vice-Chair­man of the sci­en­tif­ic coun­cil of the Uni­ver­si­ty of Tech­nol­o­gy of Com­piègne (UTC) for­mer Direc­tor of a CNRS research group (GDR) in robot­ics between 2013 and 2017. Since Jan­u­ary 2018, he has been Direc­tor of the UTC-Heudi­asyc Lab­o­ra­to­ry, cre­at­ed in 1981. This state-of-the-art lab­o­ra­to­ry hous­es, in par­tic­u­lar, the CID (Knowl­edge, Uncer­tain­ties, Data) team ded­i­cat­ed to research in arti­fi­cial intelligence. 

[UTC-Heudi­asyc is a joint research unit (UMR) between UTC and the CNRS. It employs near­ly 120 peo­ple – pro­fes­sors and lec­tur­ers, research-sci­en­tists, research engi­neers and tech­ni­cians, doc­tor­al stu­dents, post-docs, admin­is­tra­tive staff, etc. In addi­tion, there are about twen­ty Mas­ter’s lev­el trainees per year. This makes it, after UTC-Rober­val, the uni­ver­si­ty’s sec­ond largest laboratory.

“Heudi­asy­c’s sci­en­tif­ic scope has not changed much since its cre­ation, but the themes have evolved. We focus on com­put­er sci­ences in the broad­est sense, with two main branch­es: the train­ing of com­put­er engi­neers and research Mas­ter stu­dents,” explains Philippe Bon­ni­fait. This speci­fici­ty explains the very strong link between the lab and the UTC’s com­put­er sci­ence depart­ment, since the major­i­ty of the lec­tur­ers involved in the train­ing of com­put­er engi­neers (but also in the frame­work of the research Mas­ter’s degrees) report to UTC-Heudiasyc. 

UTC-Heudi­asyc is a cut­ting-edge lab­o­ra­to­ry that has seen “five of its research sci­en­tists sec­ond­ed to the pri­vate sec­tors. Among them, two are at FAIR (Face­book AI Research), the fir­m’s lab­o­ra­to­ry in Paris,” he says. A lab­o­ra­to­ry whose lec­tur­er-cum research sci­en­tists, with proven skills, pro­vide qual­i­ty train­ing that is high­ly appre­ci­at­ed by stu­dents. The proof? “There are more than 700 matric­u­lat­ed com­put­er engi­neer­ing stu­dents. This is not insignif­i­cant. It is a high­ly val­ued gen­er­al­ist train­ing that allows our stu­dents to eas­i­ly adapt to the very rapid evo­lu­tion of tech­nolo­gies in the sec­tor. For Mas­ter’s degree stu­dents, the train­ing is more spe­cial­ized, clos­er to what we do in research, in short. Cur­rent­ly, Heudi­asyc has 55 PhD stu­dents, which proves the inter­est, in the eyes of the stu­dents, in the research themes tack­led by the lab­o­ra­to­ry,” he emphasises. 

What are the major lines of research under­tak­en at UTC-Heudi­asyc? “There are four pri­or­i­ty themes: com­put­er sci­ence, arti­fi­cial intel­li­gence (AI), a key word that has been around for 20 years, when it was not yet fash­ion­able, and final­ly automa­tion and robot­ics. All our teacher-researchers work in these dis­ci­plines,” says Philippe Bonnifait. 

Since the restruc­tur­ing of the lab in Jan­u­ary 2018, three teams have been work­ing on these themes: the CID team (Knowl­edge, Uncer­tain­ties, Data), the SCOP team (Safe­ty, Com­mu­ni­ca­tion, Opti­miza­tion) and, last­ly, the SyRI team (Robot­ic Sys­tems in Inter­ac­tion). “The first team is ded­i­cat­ed to what we do in the foun­da­tion arti­fi­cial intel­li­gence stud­ies, under­stand­ing that not all areas of arti­fi­cial intel­li­gence are addressed. We can men­tion machine learn­ing, inter­ac­tive learn­ing, rec­om­men­da­tion sys­tems, etc.” The sec­ond works in par­tic­u­lar on the prob­lems of sched­ul­ing, net­works but also — and this is a nov­el­ty in the lab­o­ra­to­ry — on safe sys­tems, in oth­er words, sys­tems that are fail­safe and secure. As sys­tems com­mu­ni­cate increas­ing­ly, data exchanges are there­fore impor­tant. “Hence the major chal­lenge of secur­ing them against an attack by hack­ers, for exam­ple. Final­ly, the last theme address­es those ques­tions that revolve around the auton­o­my of machines, in par­tic­u­lar the artic­u­la­tion between robot­ics and arti­fi­cial intel­li­gence, the first being in the phys­i­cal world while the sec­ond, com­pu­ta­tion­al, vir­tu­al, is locat­ed in the “cloud” or in com­put­ers. We now speak of arti­fi­cial intel­li­gence embod­ied by robots,” he explains. 

The choice made by Heudi­asyc in the vast field of arti­fi­cial intel­li­gence applied to robot­ics? “We chose to focus our research on mobile robot­ics, par­tic­u­lar­ly those ded­i­cat­ed to trans­port and mobil­i­ty. Today, we are talk­ing about intel­li­gent (“smart”) and autonomous vehi­cles. Vehi­cles designed to dri­ve in shared envi­ron­ments. We were also one of the first labs to launch, in 1997, into UAVs,” he explains. 

A choice that allowed the lab to par­tic­i­pate, as part of the Gov­ern­ment Incen­tive Pro­gramme Invest­ments for the Future (PIA), via Labex MS2T and EQUIPEX to Robo­t­ex and enabled Philippe Bon­ni­fait to pilot every­thing relat­ed to land and air mobile robot­ics in France. A project launched in 2011 and run­ning until the end of 2021 which also had the sup­port of the Hauts-de-France Region to the tune of 3.5 mil­lion euros in sci­en­tif­ic equipment. 

This pol­i­cy choice that nat­u­ral­ly led to a large num­ber of indus­tri­al part­ner­ships, par­tic­u­lar­ly in the field of trans­port. “With Renault, for exam­ple, with­in the frame­work of Sival­ab, a joint Renault/UTC/CNRS lab­o­ra­to­ry, or a ten-year project, launched in Sep­tem­ber 2019, with the IRT Raile­ni­um con­sor­tium relat­ed to the con­cept of autonomous trains,” con­cludes Philippe Bonnifait. 

Pro­fes­sor Thier­ry Denoeux has been a senior mem­ber of the Insti­tut uni­ver­si­taire de France (IUF) since Octo­ber 2019. Direc­tor of the Labex Maîtrise des Sys­tèmes de Sys­tèmes Tech­nologiques (MS2T) with­in Heudi­asyc, a joint UTC/CNRS unit, a researcher in the Knowl­edge, Uncer­tain­ties, Data (CID) team, he is also Edi­tor-in-chief of two inter­na­tion­al jour­nals: Inter­na­tion­al Jour­nal of Approx­i­mate Rea­son­ing on uncer­tain rea­son­ing and Array, a new open-access jour­nal cov­er­ing the entire field of com­put­er sci­ence, launched in Octo­ber 2018.

As a grad­u­ate engi­neer from École des Ponts Paris­Tech (ENSP), Thier­ry Denoeux pur­sued a PhD the­sis on “the reli­a­bil­i­ty of rain fore­casts by weath­er radar” in a lab­o­ra­to­ry ded­i­cat­ed to the envi­ron­ment at ENPC. He is inter­est­ed, among oth­er things, in com­put­er sci­ence, pat­tern recog­ni­tion and image pro­cess­ing. “The goal was to process radar images to analyse and extrap­o­late the move­ment of rain cells for quan­ti­ta­tive fore­cast­ing in the very short term (one to two hours). These pre­dic­tions were used to opti­mize the man­age­ment of large sew­er sys­tems in order to lim­it flood­ing in the event of a storm,” he explains.

An inter­est that nat­u­ral­ly led him, after his PhD, to join the Lab­o­ra­toire d’In­for­ma­tique Avancée de Com­piègne (LIAC), Lyon­naise des Eaux, which has since become Suez, as a research engi­neer. He stayed there for three years and worked on Euro­pean projects with lec­tur­er cum-research sci­en­tists work­ing at UTC, in the ear­ly 1990s, when arti­fi­cial intel­li­gence (AI) was already arous­ing a lot of inter­est with the devel­op­ment of expert systems.

He joined UTC Com­pieg­ne in 1992 as a con­tract lec­tur­er-cum-research sin­guli­er sci­en­tist at Heudi­asyc before being appoint­ed as full pro­fes­sor in 1999. Sev­er­al respon­si­bil­i­ties fol­lowed suit: direc­tor of a joint lab­o­ra­to­ry with Suez, deputy direc­tor of Heudi­asyc, vice-pres­i­dent of the sci­en­tif­ic coun­cil of the UTC, sci­en­tif­ic coor­di­na­tor, before tak­ing over as direc­tor in Jan­u­ary 2019, of the Labex Maîtrise des Sys­tèmes de Sys­tèmes Tech­nologiques (MS2T) — a tenyear project — which, as part of the Gov­ern­ment Incen­tive Pro­gramme Invest­ments for the Future (PIA), run­ning until 2021. Also in Jan­u­ary 2019, he took over the man­age­ment of the SHIC¹ research fed­er­a­tion, a CNRS struc­ture ini­tial­ly group­ing togeth­er the mixed Heudi­asyc, BMBI and Rober­val units, which were recent­ly joined by the Costech unit. This fed­er­a­tion pro­vid­ed impe­tus to a new dynam­ic for inter­dis­ci­pli­nary tech­no­log­i­cal research with­in the UTC. 

At UTC-Heudi­asyc, Thier­ry Denoeux is part of the CID team in charge of arti­fi­cial intel­li­gence, struc­tured around two main research areas. The first con­cerns knowl­edge and data pro­cess­ing with themes such as knowl­edge mod­el­ling, machine learn­ing and uncer­tain­ty man­age­ment, a major chal­lenge in both arti­fi­cial intel­li­gence and sta­tis­tics. “Indeed, how can we mod­el uncer­tain­ty, rea­son and make deci­sions know­ing that we do not dis­pose of all the infor­ma­tion need­ed ?” he says. The sec­ond area of research is con­cerned with cus­tomized adap­tive sys­tems. In oth­er words, every­thing that relates to the inter­ac­tion between humans and sys­tems with the idea of design­ing sys­tems that can auto­mat­i­cal­ly and dynam­i­cal­ly adapt to the user and the con­text of use. 

Thier­ry Denoeux works main­ly on the first pri­or­i­ty theme. “I work essen­tial­ly on the mod­el­ling of uncer­tain­ties in intel­li­gent sys­tems, a theme that lies at the inter­face between arti­fi­cial intel­li­gence and sta­tis­tics. I am par­tic­u­lar­ly inter­est­ed in the the­o­ry of belief func­tions, a the­o­ry of uncer­tain­ty that allows us to rea­son and make deci­sions in the pres­ence of uncer­tain­ties. It is a gen­er­al the­o­ry, which encom­pass­es prob­a­bil­i­ty the­o­ry, and has many appli­ca­tions because uncer­tain­ties are ubiq­ui­tous. Research in this area is mul­ti­dis­ci­pli­nary and involves econ­o­mists, AI spe­cial­ists, sta­tis­ti­cians and oth­ers,” he explains. 

A field that led him, in 2010, to par­tic­i­pate in the cre­ation of a learned soci­ety Belief func­tions and Appli­ca­tions soci­ety (BFAS), an asso­ci­a­tion of which he is the Pres­i­dent. The objec­tive? In par­tic­u­lar, to pro­mote teach­ing, research, the advance­ment of knowl­edge in the field of belief func­tions and to explore the links with oth­er the­o­ries of uncer­tain­ty. Hence the launch of inter­na­tion­al con­fer­ences held every two years — the next one will be con­vened in Shang­hai in 2020 — and a the­mat­ic school for the train­ing of PhD stu­dents, the lat­est edi­tion of which was held in Octo­ber 2019 in Siena (Italy).

How­ev­er, Thier­ry Denoeux does not con­fine him­self to the the­o­ret­i­cal aspect of his research on belief func­tions, as he is also inter­est­ed in the con­crete appli­ca­tions that can derive from them. One exam­ple is auto­mat­ed postal address recog­ni­tion, which was the sub­ject of a CIFRE the­sis in part­ner­ship with Solystic, one of the world lead­ers in the pro­vi­sion of auto­mat­ed sort­ing and dis­tri­b­u­tion prepa­ra­tion solu­tions for parcels and mail. “This com­pa­ny sells machines with hand­writ­ten address recog­ni­tion soft­ware. So when the address is not rec­og­nized, the enve­lope is reject­ed and processed man­u­al­ly. The chal­lenge is to reject as few envelopes as pos­si­ble while mak­ing as few errors as pos­si­ble on those that are accept­ed. To meet these two cri­te­ria and improve machine per­for­mance, the idea was to inte­grate sev­er­al soft­ware pro­grams and com­bine the results of these sys­tems using belief func­tion the­o­ry,” he explains.

Oth­er appli­ca­tions include the work car­ried out with the French Insti­tute of Trans­port, Plan­ning and Net­work Sci­ence and Tech­nol­o­gy (IFSTTAR) and the SNCF on “Diag­no­sis of rail­way track cir­cuits”, and the ongo­ing col­lab­o­ra­tion with the Lab­o­ra­to­ry of Com­put­er Sci­ence, Infor­ma­tion Pro­cess­ing and Sys­tems (LITIS) of the Uni­ver­si­ty of Rouen on “Seg­men­ta­tion of tumours in med­ical images and prog­no­sis based on the evo­lu­tion of patient data”. The the­o­ret­i­cal cor­pus of belief func­tions is of course also of inter­est to the SyRI (Robot­ic Sys­tems in Inter­ac­tion) team, which is work­ing in par­tic­u­lar on intel­li­gent vehi­cles (IV). “One of the prob­lems in IV con­cerns per­cep­tion. IVs are full of sen­sors and the chal­lenge is to be able to process the infor­ma­tion col­lect­ed by these sen­sors to rec­og­nize objects on the road such as pedes­tri­ans, cyclists, etc. One of the prob­lems in IVs is per­cep­tion. We there­fore need to com­bine the infor­ma­tion from these dif­fer­ent sen­sors. And here again, there is a lot of uncer­tain­ty, because each sen­sor pro­vides par­tial and some­times unre­li­able infor­ma­tion about the envi­ron­ment,” adds Thier­ry Denoeux. Should we be afraid of AI? “Irra­tional fear is irrel­e­vant. How­ev­er, some appli­ca­tions of AI pose eth­i­cal prob­lems, such as gen­er­alised video sur­veil­lance with, in par­tic­u­lar, the devel­op­ment of facial recog­ni­tion. As biol­o­gists have been doing for a long time, com­put­er sci­en­tists must now be con­cerned about the eth­i­cal impli­ca­tions of their work,” he concludes.

Senior lec­tur­er, Domi­tile Lour­deaux is also a mem­ber of the Knowl­edge, Uncer­tain­ties, Data (CID) team at UTC-Heudi­asyc, a joint UTC/CNRS unit. Her research focus­es on cus­tomized adap­tive sys­tems, more specif­i­cal­ly on every­thing relat­ed to vir­tu­al real­i­ty and training.

A field that she has been work­ing on since her PhD the­sis at the École des Mines de Paris: a CIFRE the­sis fund­ed by the SNCF on “Vir­tu­al Real­i­ty and Train­ing of TGV Dri­vers”, con­duct­ed between 1998 and 2001. She pays par­tic­u­lar atten­tion to the ped­a­gog­i­cal objec­tives or “how,” she says, “vir­tu­al real­i­ty could either improve exist­ing train­ing or com­ple­ment it”.

Ini­tial­ly trained as a com­put­er sci­en­tist, she imme­di­ate­ly refused to address this ques­tion sole­ly from a tech­ni­cal point of view. “I start­ed work­ing with ergono­mists, edu­ca­tion spe­cial­ists and also end-users,” she explains. Once her PhD was com­plet­ed, she stayed for four years at the École des Mines as an asso­ciate research offi­cer, joined UTC-Com­pieg­ne in 2005 as a lec­tur­er and con­tin­ued her research based on con­crete needs, always in con­junc­tion with industrialists. 

Her first project result­ed from a meet­ing with researchers from the Nation­al Insti­tute for Indus­tri­al Envi­ron­ment and Risks (Iner­is). The objec­tive? “To ensure the train­ing of sub­con­tract­ed oper­a­tors who work on high-risk sites. This was all the more top­i­cal because of the explo­sion at the AZF plant in Toulouse, a tragedy that hap­pened only a short time ear­li­er. Explo­sion due to human errors in sub­con­tract­ing,” says Domi­tile Lour­deaux. A project that mobi­lized three PhD the­ses and obtained sig­nif­i­cant fund­ing from the Nation­al Research Agency (ANR) but also from the Region. “Usu­al­ly, vir­tu­al real­i­ty is used to train in a tech­ni­cal ges­ture or pro­ce­dure. But I want­ed to dis­tance myself from this schema. Since we are in high-risk areas, I want­ed the learn­er to be able to make, pos­si­bly make mis­takes and see the con­se­quences of his mis­takes,” she adds.

Since then, Domi­tile Lour­deaux has been work­ing on a series of new projects. This is illus­trat­ed by one on “train­ing aero­nau­ti­cal oper­a­tors in air­craft assem­bly” in part­ner­ship with STELIA Aero­space (Méaulte). Anoth­er inno­v­a­tive project? VICTEAMS (2014–2019) con­duct­ed in part­ner­ship with LIMSI in Orsay, spe­cial­ists in cog­ni­tive ergonom­ics, the Val-de-Grâce mil­i­tary med­ical school and the Paris City Fire Brigade. Project that will sure­ly be pursued. 

What is the par­tic­u­lar­i­ty of the STELIA and VICTEAMS projects? The answer is the lev­el of inte­gra­tion of arti­fi­cial intel­li­gence (AI). “Since we could­n’t afford to put dozens of devel­op­ers, like video game man­u­fac­tur­ers for exam­ple, we came up with the idea of inte­grat­ing AI to cre­ate sce­nar­ios,” she points out. “Pri­or to the STELIA project, the train­er select­ed the lev­el of the learn­er and then gave him/her the learn­ing objec­tives, for exam­ple, to work on a par­tic­u­lar safe­ty rule. The script­ing sys­tem then gen­er­at­ed learn­ing sit­u­a­tions based on these objec­tives. In the aero­nau­tics project, we want­ed to cre­ate a dynam­ic learn­er pro­file, i.e., to ensure that the sys­tem was able to detect the learn­er’s skills in order to pro­pose inter­est­ing sit­u­a­tions. We there­fore start from beliefs about his abil­i­ty to man­age, or not, the sit­u­a­tions he/she is con­front­ed with. This is called the “prox­i­mal zone” of devel­op­ment. In oth­er words, that one has skills and is capa­ble of acquir­ing new skills, close to one’s own skills. To do this, genet­ic algo­rithms and belief func­tions have been used to grad­u­al­ly extend this prox­i­mal zone,” she explains. With VICTEAMS (see box), Domi­tile Lour­deaux goes fur­ther. “I want­ed to empha­size col­lab­o­ra­tive work by cre­at­ing vir­tu­al teams. The involve­ment of dif­fer­ent play­ers and their inter­ac­tion requires an even fin­er degree of script­ing. This is notably the case for train­ing med­ical lead­ers in the man­age­ment of a mas­sive influx of injured peo­ple,” she explains. Read also on the same sub­ject Machine Learn­ing at the UTC-Heudi­asyc lab Machine Learn­ing at the UTC-Heudi­asyc lab L’UTC forme les nou­veaux tal­ents de la cyber­sécu­rité THEME : :ICTS: COMPUTER SCIENCES; AUTOMATION & CONTROL; DECISION THEORY AND APPLICATIONS L’UTC forme les nou­veaux tal­ents de la cyber­sécu­rité VR immer­sion THEME : :ICTS: COMPUTER SCIENCES; AUTOMATION & CONTROL; DECISION THEORY AND APPLICATIONS VR immersion 

Pro­fes­sor, Syl­vain Lagrue was appoint­ed to UTC-Com­pieg­ne in Sep­tem­ber 2018. In his capac­i­ty as research sci­en­tist in the Knowl­edge, Uncer­tain­ties, Data (CID) team at Heudi­asyc, a joint UTC/CNRS unit, he is work­ing on the log­i­cal rep­re­sen­ta­tion of knowl­edge and rea­son­ing, the man­age­ment of uncer­tain­ty in arti­fi­cial intel­li­gence, and deci­sion mak­ing and games.

After his DEA (cur­rent­ly the Mas­ter 2 diplo­ma) in Arti­fi­cial Intel­li­gence (AI), Syl­vain Lagrue has been work­ing on a Euro­pean project on “Tak­ing uncer­tain­ties into account for pref­er­ence mod­el­ling in geo­graph­ic infor­ma­tion sys­tems”. After com­plet­ing his the­sis, he joined the Uni­ver­si­ty of Artois in 2004 as a lec­tur­er before join­ing the UTC as a full professor.

His role with­in the CID team? “My cross­dis­ci­pli­nary pro­file allows me to work with the dif­fer­ent researchers in the team. Both in the field of “the uncer­tain” and that of “knowl­edge rep­re­sen­ta­tion”, for exam­ple,” he says. 

And how does AI fit con­crete­ly into all this? “For the pub­lic at large, AI is mag­ic made by the com­put­er. And the more mag­i­cal it is, the more AI. In oth­er words, see­ing actions made by com­put­ers that we thought were impos­si­ble,” he says.

One exam­ple among oth­ers? “Let’s con­sid­er games. When IBM’s Deep Blue defeat­ed world chess cham­pi­on Kas­parov in 1997, the gen­er­al pub­lic thought that AI was going to take every­thing in its path, and then it calmed down. The rea­son? It was noticed, after analy­sis, that what won in 1997 was the com­put­er’s com­put­ing capac­i­ty. For the gen­er­al pub­lic, this is no longer mag­ic. So it’s no longer AI,” he explains. 

But then what is AI in his mind? “It’s about mak­ing a machine rea­son when you don’t expect it to be able to do it. So there’s a whole aspect of log­ic, but also of deci­sion mak­ing. In a word, mak­ing it rea­son and make intel­li­gent choic­es,” he describes. 

This is reflect­ed in its three areas of research. The log­i­cal rep­re­sen­ta­tion of knowl­edge and rea­son­ing? “Log­ic has always — since ancient times — been a way to for­mal­ize rea­son­ing based on a cer­tain num­ber of hypothe­ses allow­ing us to draw con­clu­sions that are valid. Our objec­tive is to see this type of advanced rea­son­ing done by a machine. This can be achieved effi­cient­ly thanks to res­o­lu­tion and deduc­tion algo­rithms which, based on the hypothe­ses, ulti­mate­ly allow a machine to make deci­sions”, stress­es Syl­vain Lagrue. A skill that has led him to work on a Euro­pean project aimed at “safe­guard­ing intan­gi­ble her­itage in South-East Asia and in par­tic­u­lar the Water Pup­pets of Viet­nam, whose playlets rep­re­sent the coun­try’s his­to­ry, leg­ends, scenes from dai­ly life, etc. “All this is accom­pa­nied by music, songs and recita­tions. In terms of rich­ness, they can be com­pared to opera in Europe. So we had to rep­re­sent a lot of com­plex knowl­edge,” he says. Man­ag­ing uncer­tain­ty with AI? “If you roll a die, you don’t know which side it’s going to fall on. How­ev­er, in this case, we do have prob­a­bil­i­ties. In oth­er cas­es, we don’t even have prob­a­bil­i­ties. In the for­malisms that I use, the chal­lenge is to mod­el a sequence of “we think that such and such an action leads to this but in the oppo­site case rather to that”. In short, a much more ordi­nal mod­el­ling,” he says. Final­ly, is there an inter­est in AI games? “The advan­tage of gam­ing? It allows us to have a con­trolled uni­verse. You know what envi­ron­ment you’re in, with its pre­cise rules, whose effects you know, and you don’t have to wor­ry about the phys­i­cal aspects. It allows us to test a large num­ber of algo­rithms,” he explains. An inter­est that led him to co-direct a the­sis on “gen­er­al game play­ing”, or how to make a com­put­er play any game. “Deep Blue could only play chess, for exam­ple. In order to devel­op a pro­gram capa­ble of play­ing all games, we had to rep­re­sent all games with com­plete infor­ma­tion thanks to the Game Descrip­tion Lan­guage (GDL). Which brings us back again to the rep­re­sen­ta­tion of knowl­edge”, con­cludes Syl­vain Lagrue.

Le magazine

Avril 2024 - N°62

Faire face aux enjeux environnementaux

linkedin facebook pinterest youtube rss twitter instagram facebook-blank rss-blank linkedin-blank pinterest youtube twitter instagram