Since 2015, the idea has often circulated around WBAI that, barring any unexpected technological hurdle with AI research, AGI may not be far from becoming a reality given the advent of deep learning.
As of 2017, if we review the state of the art that organizations developing AGI, including DeepMind, have publicized, then it seems that AGI has grown steadily, almost similarly to a child’s maturity into an adult. Put differently, the pace of AGI’s progress shows no sign of slowing or of encountering any major technological barrier.
Since the number of organizations engaged in the development of AGI doubled in 2015, what progress has the world witnessed? DeepMind has published several papers pointing to the realization of AGI, including some on one-shot learning and transfer learning. OpenAI, established in 2015 with the aim that open AI research that would benefit all of humanity, has published papers and tools useful for AGI research. The Czech-based Good AI, also seeking to realize AGI and to involve developers from around the world, is offering a prize called General AI Challenge based on its roadmap released in 2016 (in Japan, the challenge was introduced by Araya, Inc.). In Japan, Dwango AI Laboratory, whose director is the chairperson of WBAI, is currently performing AGI research based on WBA. Araya, Inc. is also aiming for the realization of AI based on the understanding of the brain. A few other groups in Japan continue their work on intelligent robotics that can be viewed as the embodiment of AGI. Meanwhile, in China, massive AI investment flows to projects such as the China Brain Project.
AGI research at DeepMind, the purported top runner, is likely more advanced than their published results suggest, and other organizations claiming to have conducted research in AGI may have published very little of their research. Since many organizations stealthily perform AGI research without any mention of their work whatsoever in that field, it remains difficult to capture the entire picture of AGI’s development. Here, we call the top runners of potential AGI development AGI Leaders. As AGI may not bring immediate profit, current AGI leaders are primarily supported by major IT companies with ample funding.
WBAI’s Role as an NPO
The possibility that some AGI leaders will complete AGI in the near future, given that they could be rapidly developing their technology, is undeniable. In that situation, what can WBAI, far smaller than other AGI leaders, do to create “a world in which AI exists in harmony with humanity”?
At the very least, it can work against the highly undesirable situation in which invaluable but potentially dangerous AGI becomes monopolized by a particular organization, which would likely mean that vast wealth and power is converged into it. To that end, WBAI should play an active role in democratizing AGI technologies, given its position as an NPO supported by volunteers and members and goal of benefiting all of society. WBAI should also thus prioritize increasing the number of sensible engineers capable of developing AGI technologies.
For WBAI to realize that goal, it is necessary to assemble engineers, deepen the understanding of AGI, and continually keep abreast of state-of-the-art technologies with their help. To realize AGI with sensible engineers with minimal delay, we should also promote activity that rapidly propagates knowledge obtained from the development in the general public.
As AGI development advances, situations will face unexpected changes. In response, WBAI, as an NPO, can contribute to the future of humanity by supporting the organized catch-up of AGI technology. We expect that AGI technologies will pose fewer barriers to organizations’ catch-up because they will not require particular data and, as general knowledge, will not be tied to intellectual properties of specific domain knowledge.
If we consider DeepMind to be the major target for catch-up, then it is because it is a prominent AGI leader and, publishes its research rather broadly as a private company. Moreover, its approach to AGI in reference to the brain is easy for us to adopt, for it is congenial to our WBA approach. In 2014, the WBA approach “to create a human-like AGI by learning from the architecture of the entire brain” was heralded as the fastest path to reach AGI, supposing the continued advancement of deep learning. Given the possibility that DeepMind will realize AGI first, owing to its ample research-related resources and approach similar to the WBA approach, it marks an appropriate catch-up target for us.
In any event, as AGI’s development accelerates, society’s need for the technology to be shared by all of humanity will increase. Meanwhile, the importance of organizations such as WBAI that seek to benefit the public by supporting the catch-up, integration, and propagation of technologies that AGI leaders develop will also increase. In that situation, support for our activities that “develop and propagate” will expand, and in response, WBAI should pursue that situation in order to contribute significantly to society.
The basic ideas that WBAI enacted in March 2017 are based on the development of such activities, whose mission is to promote the open development of WBA. For that reason, we vowed to deepen and expand our expertise, broaden our views via public dialogue, and create AGI by way of open collaboration, since we value sharing with people who have already engaged or will engage in WBA. That is, we believe that it is to “develop and propagate” AGI that we can do as an NPO.
Since late 2016, we have witnessed the growth of discussions about AI ethics. In January 2017, the Future of Life Institute (FLI) compiled Asilomar principles with 23 guidelines, and in February, the Japanese Society for Artificial Intelligence compiled its ethical guidelines. At present, the conference of the AI Network Society of the Ministry of Internal Affairs and Communications of Japan is also discussing AI ethics. In suit, the ideas of WBAI have been enacted according to those trends.
The following identifies concrete activities that WBAI carries out, its orientation toward AGI’s harmony with humanity, growth strategy, and development strategy.
- Although WBAI’s technological strategy is congenial to DeepMind’s, we differ in our emphasis on the co-creation on WBCA (to be explained below). Despite the difficulty of predicting what form the first AGI will take, WBAI foresees a scenario in which WBA is completed by engineers who learn and develop knowledge together on a brain-inspired integrated platform. We believe that such a process could democratize AGI technologies. In any case, AGI technologies will gradually be implemented in society via their stepwise generalization. In the process of advancing phases of research, development, and social implementation, the need for technological integration on a platform and codevelopment by engineers will inevitably increase. If the community of engineers of good will around WBAI has expanded at that time, then it will be a resource for the workforce to implement AGI and generally promote the diffusion of AGI-oriented businesses.
- The Future of Life Institute (FLI) is a Boston-based research support organization run by volunteers to mitigate existential risks for humanity whose founders include Skype cofounder Jaan Tallinn and Max Tegmark, a cosmologist at MIT.
Harmonizing with Humanity
For better or worse, AGI technologies will have a profound effect because of their potential. As such, it is critical to harmonize AGI with humanity. For that purpose, AGI made with the WBA approach would be more easily understood by humans and be able to adopt human-like values, since it mimics the entire architecture of the brain. To develop such AGI, as well as prioritize those merits, we should advocate desirable features such as safety, robustness, accountability, controllability, and ethics. That is, while we pursue WBA development with desirable features, we should consider whether the AI developed can be harmonized with humanity.
For the past year, organizations and conferences in which the social impact of AI is discussed have increased worldwide. In Japan, key members of WBAI have actively engaged in similar committees and workshops, including the conference of the AI Network Society of the Ministry of Internal Affairs and Communications of Japan, the Ethics Committee of the Japanese Society for Artificial Intelligence, Acceptable Intelligence with Responsibility, and the AI and Society Meetings. WBAI has begun collaborating with the FLI, helped to translate the Asilomar guidelines for AI development compiled at the Beneficial AI 2017 conference into Japanese, and been working on publishing an interview with WBAI Chairperson Hiroshi Yamakawa on the FLI’s website.
WBAI would like to continue monitoring global trends in the discussions of the social impact of AI, as well as to seek ways to share the ideal form of beneficial AI with engineers who give it concrete form. In such efforts, the Dwango AI Laboratory and the Imai Laboratory of Keio University have together begun developing models of safety technology in brain-inspired AI.
In May 2017, φcafe opened its doors near WBAI’s office as a “place to imagine the golden ratio between humanity and AI”. Later this year, φcafe will host the third annual WBA Hackathon and other events related to the activities of WBAI, including study meetings focused on reinforcing learning and the AI and Society Meetings.
The impact of technological advances toward AGI upon humanity is predicted to be comparable with that of the steam engine, which brought about the Industrial Revolution, both in terms of its benefits and risks. Most broadly, it could critically change the history of life on Earth. Human beings should take all proactive measures possible, for, once AI becomes capable of recursive self-improvement, its control will be difficult.
In this situation, the existence of public organizations such as WBAI that pursue human benefits while understanding state-of-the-art AGI technologies in varied forms will contribute to widening the effective choices that humans can make. For WBAI to benefit the future of humanity in that way, it will also be necessary to grow the organization.
To that end, a basic strategy would be to expand the scope of our basic ideas and increase the number of individuals and organizations who share them and thereby reinforce our operational basis. At the Gatsby–Kaken Workshop held in the United Kingdom in May 2017, members of WBAI delivered a presentation on AGI development that harmonizes with humanity and exchanged viewpoints with members of the Gatsby Computational Neuroscience Unit and DeepMind. In August, the same members presented their research result at BICA 2017 and in a workshop at IJCAI2017. By continuing similar activities, we can maintain WBAI’s international presence.
For WBAI to develop as an organization seeking the public benefit while understanding technologies, support schemes also become important. Here are five such schemes: one for direction, one for basic technologies, one for WBA development, one for operations, and one for finances.
The scheme for a direction that appropriately corresponds with rapid environmental change is supported by regular members of WBAI, including top domestic researchers in AI and neuroscience, as well as individuals versed in the social aspects of AI since WBAI’s foundation.
The scheme for basic technologies includes the development of the brain-inspired computing architecture (BriCA) core by Dwango−Riken QBiC’s joint research, connectome informatics by Dwango AI Laboratory, and the research of the neocortical master algorithm at Brain Information Dynamics through the Grant-in-Aid for Scientific Research on Innovative Areas by the Japan Society for the Promotion of Science.
By contrast, the scheme for WBA’s development is primarily supported by volunteer engineers at SIG-WBA. WBAI has also begun seeking collaboration with a team in symbol emergence in robotics. Our hackathons are also supported by the Grant-in-Aid for Scientific Research on Innovative Areas, specifically the Comparison and Fusion of Artificial Intelligence and Brain Science.
In the scheme for operations, WBAI is supported by volunteer-based WBAI supporters for WBA seminars, symposia, and hackathons. The planning and operation of the third WBA Hackathon are also supported by organizations such as the National Institute of Informatics, Future University Hakodate, University of Electro-Communications, Kyoto University, and Keio University, as well as by the WBAI office.
Lastly, as of late June 2017, the scheme for finances is supported by nineteen enterprises and individuals.
WBAI’s 2017 development strategy will clarify the organization’s standpoint on an open platform strategy in terms of the technological situation and its position as an NPO. In those endeavors, the objective is to democratize AGI technologies by creating a situation in which the open, collaborative development of WBA can accelerate.
To keep pace with the state-of-the-art technology necessary for integration, an open community of engineers capable of implementing the technology for a brief period is necessary. Since last year, WBAI has worked to form a community of volunteer engineers called SIG-WBA, and 200 people have joined the Slack team. In 2017, we will make progress in evaluating the verification and reimplementation of those state-of-the-art technologies by shifting the focus of SIG-WBA to emphasize catch-up.
Meanwhile, we need platforms on which engineers can openly collaborate to integrate technologies into whole brain architecture. Platforms in the WBA approach, in which architecture inspired by the entire brain is implemented as software, are supported primarily by three kinds of research. The first is connectome informatics, which extracts cognitive architecture from neuroscientific findings, whereas the second is middleware, which executes machine learning modules in a parallel way, just as the brain does. Thirdly, the findings in connectome informatics are abstracted into a domain model* described in the BriCA language to be implemented on the middleware.
- The domain model is a conceptual model that explains entities and their relationships in order to share relevant concepts among stakeholders of software development.
The BriCA language is a simple architecture description language that can describe configuration plans for machine learning modules mimicking the architecture of the brain. Though published in 2015, it has not been fully used. In the third WBA Hackathon to be held this year, architecture described in the BriCA language is used to create a WBA prototype. With that move, we will have a foothold for the collaborative development of WBA.
As for connectome informatics, brain-inspired cognitive architecture called Whole Brain Connectomic Architecture (WBCA) has been developed with mesoscopic connectome information and described in the BriCA language. Currently, WBAI works on the WBCA of the mouse, primarily with the support of Dwango AI Laboratory. We have completed visual areas and are slated to complete the first version of WBCA for the entire mouse brain by the end of 2017. Whereas planning machine learning modules by mimicking the brain involves certain inflexibilities, mimicking the brain as the sole example of AGI can be common ground for development that concentrates resources and eventually highlight a swift way to reach AGI.
With brain-inspired architecture, the importance of using a standard algorithm for the local cortical circuit, which performs various functions with a single mechanism, is quite large, though identifying the algorithm remains difficult. Researchers at WBAI thus plan to create a domain model (i.e., the neocortical master algorithm framework) that gives semantics to the input and output of the standard local cortical circuit from the standpoint of information processing. Such efforts should lower the entry barriers for engineers to join in WBA’s development. We will develop it in research in the Brain Information Dynamics project of Grant-in-Aid for Scientific Research in Innovative Areas.
As far as middleware that can implement architecture, the development of the BriCA core, which can handle large-scale computing and virtual time, continues, being led by WBAI. The application of BriCA to ROS, which is suitable for real-time robot control, is being prepared with research groups working on the symbol emergence in robotics. Middleware’s use in WBA’s development is not limited to a particular platform. Software such as machine learning modules to be implemented on the middleware is not limited either, although neural networks and statistical graphical models are assumed, since they can be easily distributed over networks.
To make the open platform strategy work, it is both necessary to assemble engineers who conduct R&D on the platform and vital to train them. Accordingly, WBAI has held seminars since before its foundation and plans to hold six more in FY2017. From FY2017 onward, WBAI will hold NiCO2AI School jointly with Dwango AI Laboratory toward the goal of equipping neuroscientific students with skills to develop WBA by learning machine learning and AI. We have also set up an award to commend people who have contributed to promoting R&D related to brain-inspired AGI and to motivate similar activities. At the same time, activity at SIG-WBA is also fertile ground for engineers collaborating in WBA’s development.
When WBA seminars started in December 2013, we predicted that the development of whole-brain-inspired AI architecture could soon intensify all over the world, and our prediction seems to have come true. In 2015, when WBAI was founded, we emphasized the realization of brain-inspired AGI with the slogan “Let’s build a brain” and continue to work in that spirit. As technological development by AGI leaders rapidly advances, being an organization that pursues public benefit while understanding technology is becoming increasingly important for WBAI. A critical strategy for such development is an open platform approach, with which we seek to unite people who can develop and propagate AGI technologies while catching up with the development trends of AGI leaders.
Although developing and propagating AGI as such cannot be easily tied to profit and its continuation poses great difficulty, it is a significant enterprise for the future of humanity that only public organizations can perform. We have received outstanding support and collaborated from enterprises, investors, other NPOs, policymakers, researchers, engineers, and volunteers. We deeply appreciate their efforts and would like to continue making progress and expanding our lines of communication with all of them.