3 avril 2020 | International, Aérospatial

A la recherche de l'IA hybride

Vice-président recherche, technologie, innovation du groupe Thales, David Sadek intervenait jeudi 24 octobre à Toulouse au Forum Innovation IA, un événement organisé par Sciences et Avenir pour le groupe Challenges.

Explicabilité, validité, intégrité... " Pour une IA de confiance " : c'était l'intitulé de l'intervention de David Sadek, le vice-président recherche, technologie, innovation du groupe Thales. "Nous aimerions tous pouvoir toujours faire confiance à tous nos outils technologiques. Mais en réalité tout dépend du contexte. Dans le cas des outils reposant sur l'IA, si Netflix me recommande un film que je trouve détestable, ce n'est pas bien grave, on s'en remettra. Mais dans le cas des systèmes critiques tels que l'on en conçoit régulièrement chez Thales pour des applications défense, espace et aéronautique, la question de la confiance est primordiale".

David Sadek énonce trois impératifs, qui sont autant d'axes stratégiques de la feuille de route IA chez Thales, pour aller vers cette " IA de confiance ". Pour qu'une application de l'IA soi digne de confiance, il faut d'abord qu'elle soit capable d'expliquer. Si elle se contente de proposer ou décider sans être capable de dire pourquoi et comment elle est arrivée à cette proposition ou décision, la défiance s'installera. Elle doit donc être capable de répondre à la question : " Pourquoi ? ".

"Si par exemple un copilote IA recommande au pilote humain de virer à 45°, et que ce dernier demande " pourquoi ", le copilote doit pouvoir répondre par exemple : " parce qu'il y a un problème météo (ou une menace ...) droit devant ". Il ne s'agit pas de simplement tracer la décision, de pouvoir indiquer à ses concepteurs quels neurones, quelles couches de neurones, ont fait pencher la balance dans un sens ou un autre, mais bien d'expliquer dans une langue compréhensible immédiatement par le pilote". Ce qui indique qu'à l'aspect intelligence artificielle s'ajoute celui de " l'interaction homme-machine ".

IA connexionniste et IA symbolique

Cela implique, en restant sur cet exemple, une capacité de compréhension et de génération du langage naturel. Mieux, le pilote appréciera de pouvoir dialoguer dans un langage mixte, reposant sur l'oral et le geste (pour désigner un point sur une carte, etc.). Deuxième impératif selon David Sadek : il faut pouvoir démontrer la validité d'une application de l'IA. C'est-à-dire sa conformité aux spécifications. "Le système développé doit faire tout ce l'on attend de lui et rien que ce que l'on attend de lui. Ce qui suppose que l'on sache spécifier très proprement". On connait l'exemple d'un jeu de bataille navale, un " serious game " destiné à former des officiers de la marine, qui gagnait à tous les coups, parce qu'il sacrifiait systématiquement chaque vaisseau touché, pour ne pas ralentir la flotte.

Aucune règle ne prévoyait, n'interdisait ce cas de figure, passablement inacceptable. Troisième impératif pour une " IA de confiance ", selon David Sadek, la responsabilité des systèmes reposant sur l'IA. C'est-à-dire leur conformité aux cadres légaux, réglementaires et moraux. On connait le tendon d'Achille des réseaux de neurones : ce sont des boîtes noires, on voit ce qui entre et ce qui sort, mais on ne sait pas (trop) ce qu'il se passe à l'intérieur. C'est pourquoi David Sadek insiste sur le fait qu'il faut s'intéresser aussi à " l'autre IA ", l'IA symbolique, celle qui repose sur des règles et des raisonnements et qui s'oppose à l'IA connexionniste des réseaux de neurones. Cette IA symbolique, qui fait moins les gros titres, qui n'a pas connu les progrès fulgurants de l'IA des réseaux de neurones, est plus à même de répondre à ces trois exigences : expliquer, valider, responsabilité. C'est pourquoi la réponse à cette triple exigence passe sans doute, entre autres, par " l'IA hybride ", une IA tirant parti de ces deux branches de l'intelligence artificielle.

Par Pierre Vandeginste

https://www.sciencesetavenir.fr/high-tech/intelligence-artificielle/ia-pvdg_138526

Sur le même sujet

  • Study finds these gaps in Army’s small unit counter-drone capabilities

    6 juillet 2018 | International, Aérospatial

    Study finds these gaps in Army’s small unit counter-drone capabilities

    Army units at and below the battalion level are unprepared to defeat aerial drones and current plans can't keep up with rapidly evolving technology, according to a recent study. Back in 2016, the Army Research Office asked an outside organization, The National Academies of Science, Engineering and Medicine, to evaluate their counter drone capabilities for battalion and below operations. The report they published earlier this year notes some significant gaps and threats to soldiers with this technology. “Contrary to the past, when U.S. warfighters may have found (improvised explosive devices), now the IEDs will find our warfighters,” according to the report. While the Army and Marine Corps, which also included representatives in the study, are throwing resources at the small drone problem, they are not keeping pace with the threat. “Army time frames are significantly out of sync with the rapidly advancing performance capabilities of individual (small Unmanned Aerial Systems) and teams of sUASs,” according to the report. The report noted that most of the service's counter drone asset work was focused on heavy vehicle platforms or on fixed sites, which leaves smaller units most likely to first encounter the threat more exposed. “Significant quantities of man-portable” counter-drone systems have been fielded, Army spokesman Maj. Chris Ophardt told Army Times in an email. The Army will continue to pursue those capabilities based on emerging threats. Based on his response, which did not include details of capabilities, the Army is pursuing other ways to defeat drones. A large portion of the study was classified, due to operational security concerns. “Future Army C-UAS systems will encompass a variety of potential platforms to include fixed, mobile, and Soldier-portable capabilities,” Ophardt wrote. But beyond the types of systems employed, what they're targeting or attacking also came under fire in the report. The Army and other branches have invested significantly in counter-drone technology, “often focusing on detecting radio frequency transmissions and GPS signals of individual sUASs. However, today's consumer and customized sUASs can increasingly operate without radio frequency (command and control) links.” Drones now available can use automated target recognition, tracking, obstacle avoidance and other software-enabled activities instead of traditional RF and GPS. Ophardt did not divulge specifics of how the Army is addressing this, but responded that the service's counter drone capabilities, “include multiple methods in order to detect, identify and defeat enemy UAS.” A new school began last month at Fort Benning, Georgia to give basic trainees familiarity with small drones. The drone school gives infantry and scouts the ability to fill out a seven-line report when they encounter a drone then relay that info to their headquarters. The students use both fixed-wing and helicopter small drones. They also learn defensive tactics such as how to use dispersal and hiding tactics to minimize casualties from drone-coordinated fires, according to an Army release. Those introductory tactics can help even brand-new soldiers start thinking about how to deal with drone threats. But, at the same time, the low-level tactics currently used for counter drone work have tried to use “kinetic effects,” basically shooting down the drone by interfering with its signals or overheating its circuits. The report noted that method isn't practical on a wide scale for large numbers of troops, especially dismounted units. That path only adds more gear from the equipment to the batteries, to an already overloaded soldier, not to mention the “cognitive load” of training and using another piece of equipment, according to the report. Ophardt responded that the Army's counter-drone strategy included “multiple methods” to detect, identify and defeat” enemy drones. The major provided a similar response when asked about Army efforts at counter-drone tactics, capabilities against swarming drones and collaboratively acting drone groups, which the report remarks will be more prevalent and sophisticated as soon as 2025. Report authors urge Army leaders to adjust their timelines for matching tech development, which are woefully inadequate for the exponential changes in software, hardware and drone capabilities. Current Army time frames consider near-term planning to run from now until 2025; mid-term planning in the 2026 to 2035 window and far-term at the 2036 to 2050. Those efforts mirror vehicle acquisition strategy timelines, not the drone arena. The report pushes for a near-term planning of one to two years, mid-term at the three- to five-year level and far term in drone tech at the six- to eight-year range. The advances are happening so quickly, authors point out, that it is “impossible to predict performance capabilities beyond eight years.” https://www.armytimes.com/news/your-army/2018/07/05/study-finds-these-gaps-in-armys-small-unit-counter-drone-capabilities

  • Contract Awards by US Department of Defense - September 11, 2019

    12 septembre 2019 | International, Aérospatial, Naval, Terrestre, C4ISR, Sécurité

    Contract Awards by US Department of Defense - September 11, 2019

    NAVY The Boeing Co., St. Louis, Missouri, is awarded a $45,803,988 firm-fixed-price delivery order (N61340-19-F-0135) against a previously issued basic ordering agreement (N00019-19-G-0002) to procure P-8A aircrew training system production concurrency upgrades for the Navy and the government of Australia. Work will be performed in St. Louis, Missouri (45%); Jacksonville, Florida (40%); Adelaide, Australia (12%); Whidbey Island, Washington (2%); and Orlando, Florida (1%), and is expected to be completed in December 2022. Fiscal 2017 aircraft procurement (Navy) funds in the amount of $37,000,969; and cooperative engagement agreement funds in the amount of $8,803,019 are being obligated on this award, $37,000,969 of which will expire at the end of the current fiscal year. The Naval Air Warfare Center Training Systems Division, Orlando, Florida, is the contracting activity. CM Construction Services Inc.,* Visalia, California, is awarded a $20,000,000 indefinite-delivery/indefinite-quantity contract for other specialty trade contractors construction alterations, renovations and repair projects at Naval Support Activity Monterey. Projects will be primarily design-bid-build (fully designed) task orders or task orders with minimal design effort (e.g. shop drawings). Projects may include, but are not limited to, alterations, repairs, and construction of electrical; mechanical; painting; engineering/design; paving (asphaltic and concrete); flooring (tile work/carpeting); roofing; structural repair; fencing; heating, ventilation and air conditioning; and fire suppression/protection system installation projects. Work will be performed in Monterey, California. The term of the contract is not to exceed 60 months, and is expected to be completed September 2024. Fiscal 2019 operation and maintenance, (Navy) (O&M, N) contract funds in the amount of $5,000 are obligated on this award and will expire at the end of the current fiscal year. Future task orders will be primarily funded by O&M, N. This contract was competitively procured via the Navy Electronic Commerce Online website with 18 proposals received. The Naval Facilities Engineering Command Southwest, San Diego, California, is the contracting activity (N62473-19-D-2608). Veterans Northwest Construction LLC,* Seattle, Washington, is awarded a $12,277,000 firm-fixed-price task order N44255-19-F-4417 under a multiple award construction contract (N44255-17-D-4015) for a special project (repair railroad tracks), Naval Base Kitsap, Bremerton, Washington. The work to be performed includes repair to three railroad bridges and track modifications. Work will be performed in Shelton and Bremerton, Washington, and is expected to be completed by June 2021. Fiscal 2019 operations and maintenance, (Navy) contract funds for $12,277,000 are obligated on this award and expire at the end of the current fiscal year. Two proposals were received for this task order. The Naval Facilities Engineering Command, Northwest, Silverdale, Washington, is the contracting activity (N44255-17-D-4015). ARMY Eastman Aggregate Enterprises LLC,* Lake Worth, Florida, was awarded a $15,949,855 firm-fixed-price contract for nourish critically eroded shoreline along Miami-Dade Beach. Bids were solicited via the internet with seven received. Work will be performed in Miami, Florida, with an estimated completion date of July 17, 2020. Fiscal 2019 civil construction funds in the amount of $15,949,855 were obligated at the time of the award. U.S. Army Corps of Engineers, Jacksonville, Florida, is the contracting activity (W912EP-19-C-0025). Golden Wolf Ewing Cole JV, Huntington, Maryland (W912DY-19-D-0020); HKS WSP JV, Dallas, Texas (W912DY-19-D-0021); and Rogers, Lovelock & Fritz Inc., Orlando, Florida (W912DY-19-D-0022), will compete for each order of the $9,900,000 firm-fixed-price contract for the procurement of specialized medical facilities architect-engineering services. Bids were solicited via the internet with 17 received. Work locations and funding will be determined with each order, with an estimated completion date of Sept. 10, 2021. U.S. Army Corps of Engineers, Huntsville, Alabama, is the contracting activity. *Small Business https://www.defense.gov/Newsroom/Contracts/Contract/Article/1957779/source/GovDelivery/

  • Podcast: What Could Go Wrong In 2022

    14 janvier 2022 | International, Aérospatial, Naval, Terrestre, C4ISR, Sécurité

    Podcast: What Could Go Wrong In 2022

    As the aerospace industry learns to live with COVID, it faces other big challenges from a stressed supply chain to geopolitical disruptions. Listen in as our editors discuss.

Toutes les nouvelles