1 Not for Public Release until Approved by the 2 Senate Armed Services Committee 3 4 5 6 Statement of Dr Peter Highnam 7 Deputy Director Defense Advanced Research Projects Agency DARPA 8 9 Department of Defense Artificial Intelligence Initiatives 10 11 Before the 12 Emerging Threats and Capabilities Subcommittee 13 Committee on Armed Services 14 United States Senate 15 16 17 March 12 2019 18 19 20 21 22 23 24 25 26 27 28 29 1 30 DARPA’s Seminal Role in the Field of Artificial Intelligence 31 32 Seventy years ago when early electronic computers ran on vacuum tubes and filled entire rooms 33 researchers already were striving to enable machines to think as people do Only a few years after its 34 start in 1958 DARPA began playing a central role in realizing this ambition by laying some of the 35 groundwork for the field of artificial intelligence AI Early work in AI emphasized handcrafted 36 knowledge and computer scientists constructed so-called expert systems that captured the rules that 37 the system could then apply to situations of interest Such “first wave” AI technologies were quite 38 successful – tax preparation software is a good example of an expert system – but the need to handcraft 39 rules is costly and time-consuming and therefore limits the applicability of rules-based AI 40 technologies 41 The past few years have seen an explosion of interest in a sub-field of AI dubbed “machine learning” 42 that applies statistical and probabilistic methods to large data sets to create generalized representations 43 that can be applied to future samples Foremost among these approaches are deep learning artificial 44 neural networks trained to perform a variety of classification and prediction tasks when adequate 45 historical data is available Therein lies the rub however as the task of collecting labelling and 46 vetting data on which to train such “second wave” AI techniques is prohibitively costly and time- 47 consuming 48 DARPA envisions a future in which machines are more than just tools that execute human- 49 programmed rules or generalize from human-curated data sets Rather the machines DARPA 50 envisions will function more as colleagues than as tools Towards this end DARPA is focusing its 51 investments on a “third wave” of AI technologies that brings forth machines that can reason in 52 context Incorporating these technologies in military systems that collaborate with warfighters will 53 facilitate better decisions in complex time-critical battlefield environments enable a shared 54 understanding of massive incomplete and contradictory information and empower unmanned 55 systems to perform critical missions safely and with high degrees of autonomy 56 Today DARPA is funding more than 24 programs exploring ways to advance the state of the art in 57 AI pushing beyond second wave machine learning towards contextual reasoning capabilities This 58 is in addition to more than 55 active programs that are leveraging machine learning or AI 59 technologies in some capacity–from managing the electromagnetic spectrum to detecting and 60 patching cyber vulnerabilities 2 61 This level of investment has been years in the making and will define scientific and technical 62 exploration as well as resulting military capabilities for decades to come 63 Current Programs 64 DARPA’s Lifelong Learning Machines L2M program is exploring ways to enable machines to 65 learn while doing without catastrophic forgetting Such a capability would enable systems to improve 66 on the fly recover from surprises and keep them from drifting out of sync with the world 67 First announced in 2017 L2M research teams are developing complete systems and their components 68 as well as exploring learning mechanisms in biological organisms with the goal of translating them 69 into computational processes Discoveries in both technical areas are expected to generate new 70 methodologies that will allow AI systems to learn and improve during tasks apply previous skills and 71 knowledge to new situations incorporate innate system limits and enhance safety in automated 72 assignments While the program is still in its early stages L2M researchers already have identified 73 and solved challenges associated with building and training a self-reproducing neural network 74 DARPA is also currently running a program called Explainable AI or XAI to develop new machine- 75 learning architectures that can produce accurate explanations of their decisions in a form that makes 76 sense to humans As AI algorithms become more widely used reasonable self-explanation will help 77 users understand how these systems work and how much to trust them in various situations XAI 78 specifically aims to create a suite of machine learning techniques that produce explainable models – 79 while maintaining a high level of prediction accuracy so human users understand appropriately trust 80 and effectively manage the emerging generation of artificially intelligent partners Enabling 81 computing systems in this manner is critical because sensor information and communication systems 82 generate data at rates beyond what humans can assimilate understand and act upon 83 The real breakthrough for artificial intelligence however will not come until researchers figure out 84 a way for machines to learn or otherwise acquire common sense Without common sense AI systems 85 will be powerful but limited tools that require human inputs to function With common sense an AI 86 could become a partner in problem solving Common sense knowledge is so pervasive in our lives 87 that it can be hard to recognize For example in conflict and warzone situations people tend to make 88 snap decisions about the cause of the problem and ignore evidence that does not support their point 89 of view To act as a valued partner in such situations the AI system will need sufficient common 90 sense to know when to speak and what to say which will require that it have a good idea of what each 3 91 person knows Interrupting to state the obvious would quickly result in its deactivation particularly 92 under stressful conditions 93 In order to find answers to the common sense problem DARPA launched in October of last year the 94 Machine Common Sense MCS program which will explore recent advances in cognitive 95 understanding natural language processing deep learning and other areas of AI research MCS is 96 pursuing two approaches for developing and evaluating different machine common sense services 97 The first approach seeks to create computational models that learn from experience and mimic the 98 core domains of cognition as defined by developmental psychology This includes the domains 99 of objects intuitive physics places spatial navigation and agents intentional actors Researchers 100 will develop systems that think and learn as humans do in the very early stages of development 101 leveraging advances in the field of cognitive development to provide empirical and theoretical 102 guidance 103 To assess the progress and success of the first strategy’s computational models researchers will 104 explore developmental psychology research studies and literature to create evaluation criteria 105 DARPA will use the resulting set of cognitive development milestones to determine how well the 106 models are able to learn against three levels of performance prediction expectation experience 107 learning and problem solving 108 The second MCS approach will construct a common sense knowledge repository capable of 109 answering natural language and image-based queries about common sense phenomena by reading 110 from the Web DARPA expects that researchers will use a combination of manual construction 111 information extraction machine learning crowdsourcing techniques and other computational 112 approaches to develop the repository The resulting capability will be measured against the Allen 113 Institute for Artificial Intelligence AI2 Common sense benchmark tests which are constructed 114 through an extensive crowdsourcing process to represent and measure the broad common sense 115 knowledge of an average adult 116 AI Next Campaign 117 DARPA announced in September 2018 a multi-year investment of more than $2 billion in new and 118 existing programs called the “AI Next” campaign Campaign key areas include providing robust 119 foundations for second wave technologies aggressively applying second wave AI technologies into 120 appropriate systems and exploring and creating third wave AI science and technologies 4 121 AI Next builds on DARPA‘s five decades of AI technology creation to define and to shape the future 122 always with the Department’s hardest problems in mind Accordingly DARPA will create powerful 123 capabilities for the DoD by attending specifically to the following areas 124 New Capabilities AI technologies are applied routinely to enable DARPA R D projects including 125 more than 60 ongoing programs such as the Electronic Resurgence Initiative and other programs 126 related to real-time analysis of sophisticated cyber attacks detection of fraudulent imagery 127 construction of dynamic kill-chains for all-domain warfare human language technologies multi- 128 modality automatic target recognition biomedical advances and control of prosthetic limbs DARPA 129 will advance AI technologies to enable automation of critical Department business processes One 130 such process is the lengthy accreditation of software systems prior to operational deployment 131 Automating this accreditation process with known AI and other technologies now appears possible 132 Robust AI AI technologies have demonstrated great value to missions as diverse as space-based 133 imagery analysis cyber attack warning supply chain logistics and analysis of microbiologic systems 134 At the same time the failure modes of AI technologies are poorly understood DARPA is working to 135 address this shortfall with focused R D both analytic and empirical DARPA’s success is essential 136 for the Department to deploy AI technologies particularly to the tactical edge where reliable 137 performance is required 138 Adversarial AI The most powerful AI tool today is machine learning Machine learning systems are 139 easily duped by changes to inputs that would never fool a human The data used to train such systems 140 can be corrupted and the software itself is vulnerable to cyber attack These areas and more must 141 be addressed at scale as more AI-enabled systems are operationally deployed 142 High Performance AI Computer performance increases over the last decade have enabled the 143 success of machine learning in combination with large data sets and software libraries More 144 performance at lower electrical power is essential to allow both data center and tactical deployments 145 DARPA has demonstrated analog processing of AI algorithms with 1000 times speedup and 1000 146 times power efficiency over state-of-the-art digital processors and is researching AI-specific 147 hardware designs DARPA is also attacking the current inefficiency of machine learning by 148 researching methods to drastically reduce requirements for labeled training data 149 Next Generation AI The machine learning algorithms that enable face recognition and self-driving 150 vehicles were invented over 20 years ago DARPA has taken the lead in pioneering research to 5 151 develop the next generation of AI algorithms which will transform computers from tools into 152 problem-solving partners DARPA research aims to enable AI systems to explain their actions and 153 to acquire and reason with common sense knowledge DARPA R D produced the first AI successes 154 such as expert systems and search and more recently has advanced machine learning tools and 155 hardware 156 In addition to new and ongoing DARPA research a key component of the AI Next campaign will be 157 DARPA’s Artificial Intelligence Exploration AIE program first announced in July 2018 AIE 158 constitutes a series of high-risk high payoff projects where researchers work to establish the 159 feasibility of new AI concepts within 18 months of award Leveraging streamlined contracting 160 procedures and funding mechanisms enables these efforts to move from proposal to project kick-off 161 within 3 months of an opportunity announcement 162 Conclusion 163 Over its 60-year history DARPA has made significant investments in the creation and advancement 164 of artificial intelligence technologies that have produced game-changing capabilities for the 165 Department of Defense and beyond DARPA’s AI Next effort is simply a continuing part of its 166 historic investment in the exploration and advancement of AI technologies 167 168 Current R D investment around the world is largely focused on second wave AI or machine learning 169 which is very good in finding patterns in voice and imagery and has many commercial applications 170 The difference is in the United States DARPA is aggressively pursuing programs that will make 171 second wave AI more robust for defense and security applications all while helping realize the third 172 wave of AI or contextual reasoning DARPA has unique access to the United States’ world-class 173 science and technology community comprised of leading universities government labs and industry 174 partners – this mix cannot be found or replicated anywhere else in the world Marshalling those unique 175 resources the Agency’s third wave research efforts will forge new theories and methods that will 176 make it possible for machines to adapt contextually to changing situations advancing computers from 177 tools to true collaborative partners Going forward the agency will be fearless about exploring these 178 new technologies and their capabilities – DARPA’s core function – pushing critical frontiers ahead 179 of our nation’s adversaries 6