Our activities in FY 2020 (from April 2020 to March 2021 — the sixth year) include educational and R&D businesses, implemented according to our policy for FY 2020.

WBAI’s activities have been conducted with non-paid volunteers, except for a paid part-time worker at the secretariat.

Educational Business

The goal of our educational business is to help people who conduct research on the WBA approach on a long-term basis. In FY2020, WBAI held two WBA seminars and the fifth WBA symposium.  We also held our first “WBA Lecture” as a new series of events related to the WBA approach.  With financial consideration, we decided to charge non-student participants of WBA seminars from FY2020.  We also decided to hold the event online due to the COVID-19 pandemic. Videos of the paid WBA seminars are distributed on Vimeo for a fee, and those from free events are distributed on YouTube.

WBA Seminars

In FY2020, WBAI held two seminars (in Japanese) with the following themes and speakers.

• 30th Seminar: April 20, 2020, General AI and symbiotic interaction
  with Takufumi Yanagisawa (Osaka Univ.) and Michita Imai (Keio Univ.)
• 31st Seminar: October 23, 2020, Cognitive Models of Sociality
  with Masatoshi Yoshida (Hokkaido Univ.) and Shigeru Taguchi (Hokkaido Univ.)

The Fifth WBA Symposium

The symposium was held (in Japanese) online on October 19, 2020 with the theme of “The Ridge of Intelligence Research Coming into Sight” with speakers Ryutaro Ichise (NII), Yutaka Matsuo (University of Tokyo), and Kenji Doya (OIST).  There were also presentations by Hiroshi Yamakawa and Naoya Arakawa (both WBAI).

It was organized by WBAI and supported by the Japanese Society for Artificial Intelligence, Grant-in-Aid for Scientific Research on Innovative Areas “Comparison and Fusion of Artificial Intelligence and Brain Science,” Program for Promoting Researches on the Supercomputer Fugaku “Human-scale whole brain simulation with connectome analysis and structure-function estimation,” and MEXT “Brain information dynamics underlying multi-area interconnectivity and parallel processing.”

In the symposium, the Award for Meritorious Service was presented to Takahiro Ikushima for his contributions to the organization of the volunteer group.  Matthew Crosby and Benjamin Beyret, who organized the Animal-AI Olympics, and Masahiro Suzuki, who developed the Joint Multimodal Variational Autoencoder (JMVAE) and  Neuro-Serket, were conferred the WBA Incentive Awards.

A Tutorial at IJCAI-20

January 8th, 2020
Ryota Kanai (Araya), Masataka Watanabe (MinD in a Device), and Hiroshi Yamanaka (WBAI) gave an online tutorial “Conscious AI: Significance and Development” at IJCAI-20 with 20 participants [1].

Working Memory Modelathon

July-September 2020
We solicited models of working memory prior to the fifth WBA Hackathon (see below).
Since the assessment methodology for biological plausibility was developed in the process of the assessment of an application, we confer the applicant the Significant Contribution Award.

The First WBA Lecture

February 7th, 2021
Theme: Decomposition of cognitive functions according to brain structure — Toward a systematic understanding of brain functions (in Japanese)
Lectureres: Hiroshi Yamanaka (WBAI) , Ayako Fukawa (WBAI) and Yoshimasa Tawatsuji (Waseda Univ.)
Online: Zoom Webinar (with 190 participants) (free admission)

The Fifth WBA Hackathon — Beyond ‘now and here’ —

May-August 2021
The theme is to create WBA agents to solve tasks with working memory.  We prepared for it in 2020 and it is being held online in collaboration with Cerenaut, WBAI’s partner.
(⇒ the Call for Participation page)

R&D Business

The goal of our R&D business is to support research activities in the WBA approach.

Development methodology

Since FY 2018, we have been discussing the methodology for the development of brain-inspired AI.  It has taken the form of BRA-driven development, which is divided into the task of designing a brain reference architecture (BRA) and the task of developing software using the BRA, as shown in Fig. 1. The BRA consists of the Brain Information Flow (BIF [2][3]), a mesoscopic-level anatomical structure of the brain, and Hypothetical Component Diagrams (HCD) that represent hypotheses of functional computational functions consistent with the BIF.  A BIF Circuit represents a subnetwork of a mesoscopic network in the brain.  Connections are set up between Circuits as outputs from Uniform circuits, which correspond to groups of neurons of a particular cell type.  An HCD is a UML component diagram, a directed graph describing the dependencies between components, and the functions performed by the regions of interest (ROI) in the brain in a way consistent with the mesoscopic-level anatomy.  It is ‘hypothetical’ in the sense that there is no guarantee that it is consistent with what is actually going on in the brain.

Fig.1

HCDs are directly used in BRA-driven software development. For most brain regions, the description of organized functions, such as component diagrams, cannot be obtained directly from neuroscientific knowledge. Thus, we have been creating BRAs with a method called ‘the Structure-constrained Interface Decomposition (SCID) method,’ which develops HCDs in a form consistent with neuroscientific findings (mainly anatomical structures – BIFs) as described in [4].

Collaborating with local research groups (R&D promotion)

WBAI has been collaborating with other organizations to develop research infrastructure such as software to support research using the WBA approach. In FY2020, we collaborated on the following research and development activities.  In particular, we collaborated with the Grant-in-Aid for Scientific Research on Innovative Areas “Comparison and Fusion of Artificial Intelligence and Brain Science” on the use of probabilistic generative models [5][6].  We also collaborated with the Kakenhi project “Brain information dynamics underlying multi-area interconnectivity and parallel processing” on establishing a foundation for neuroinformatics (below).

Establishing a foundation for neuroinformatics

The foundation of BRA is the mesoscopic anatomy of the whole brain. We analyzed the Allen Institute’s Mouse Connectome, one of the richest mesoscopic connectomes in mammals, to extract reliable connection information.

BRA design (SCID method)

As for the BRA design with the SCID method for brain organs, we assigned hypotheses on attention to the BIF for the basal ganglia loop described since 2019 [7], proposed a hypothesis on the path integral function of the medial entorhinal cortex [8], and constructed a BRA for eye movements [9].

As for the wider function of the brain, we examined hypotheses on the general meaning of signals circulating within the neocortex [10][11].

Biological Plausibility Assessment

Biological plausibility assessment is to assess how well brain-based software reproduces the states of affairs in the brain as observed in neuroscience at a granularity down to the Uniform Circuit level.  In the WBA approach, which explores AGI in a design space close to the brain, it is necessary to establish the method of assessment and continue to evaluate software.  We have studied it while assuming BRA is used [12][13], and established a method that divides the object of assessment into the consistency of BRA with neuroscientific knowledge and the fidelity of brain-inspired software with BRA.

Machine Learning

Expanding on previous work, we studied reinforcement learning in the space of hidden macroscopic time series [14].

Activities to make AGI Beneficial

Because of the potentially large impacts of AGI on humankind, we are conducting activities to make it beneficial, safe, and democratized.

WBAI Activities and Volunteering

WBAI activities such as WBA seminars have been conducted with non-paid volunteers, except for a paid part-time worker at the secretariat.

Financial Statements for FY2020

The balance sheet and cash flow for FY2020 are presented below (Table 1 and Table 2).

The revenues were 3.60 million yen and the expenditure totaled 0.82 million yen for administration and 1.61 million yen for operation (total expenditure was 2.43 million yen to yield 1.17 million yen surplus).

The businesses of WBAI have been financially supported by sponsors or supporting members.  (As of March 2021, there were ten supporting members consisting of enterprises and individuals.)  Two supporting members at the foundation paid fees for five years in advance, part of which is incorporated in the revenue in FY2020.  The other incomes include fees from the participants of the WBA seminars and symposium.

Honoraria were paid to the speakers of the WBA Seminars.  The subcontractor/outsourcing expenses for operation includes the cost of R&D,  which was mostly carried out with the budget of exterior research institutions (notably the Grant-in-Aid for Scientific Research on Innovative Areas “Comparison and Fusion of Artificial Intelligence and Brain Science” via Riken).  Prizes/awards were paid to the awardees of the WBA Incentive Awards and the Significant Contribution Award for the Working Memory Modelathon.

The expenses for secretariat personnel (shown in Subcontracting/outsourcing Expenses) were apportioned by 50% (50/50) for operating and administrative expenses.  The remuneration was paid to an accountant office.  The rent for the office was free of charge by courtesy of Garm LLC.

Table 1: Balance Sheet (as of March 31, 2021)

Items					Amount (JPY)
Ⅰ	Assets
	1.	Current Assets
		Cash and Saving Account			9,145,734
		Total Current Assets			9,145,734
	Total Assets				9,145,734
Ⅱ	Liabilities
	1.	Current Liabilities
		Advance Received			400,000
		Withholding Taxes			6,126
		Total Current Liabilities			406,126
	Total Liabilities				406,126
Ⅲ	Net Assets
		Retained Net Assets at the Beginning of Period			7,569,416
		Net assets variation			1,170,192
	Total Net Assets				8,739,608
	Total Liabilities and Net Assets				9,145,734

Table 2: Cash Flow

Items					Amount (JPY)
Ⅰ	Recurring Revenues
	１	Fees
		Fees from Regular Members			150,000
		Fees from Supporting Members			3,080,000
		Total Fees			5,190,000
	２	Other Revenues
		Interest Income			76
		Other Income			367,963
	Total Recurring Revenues				3,598,039
Ⅱ	Ordinary Expenses
	１	Operating Expenses
		⑴	Total Personnel Expenses		162,750
		⑵	Other Expenses
			Honoraria		72,000
			Subcontractor/outsourcing Expenses		951,285
			Travel Expenses		0
			Communication		284,525
			Prizes/Awards		130,072
			Other		7,821
			Total Other Expenses		1,445,703
		Total Operating Expenses			1,608,453
	２	Administrative Expenses
		⑴	Total Personnel Expenses		0
		⑵	Other Expenses
			Subcontractor/Outsourcing Expenses		533,500
			Remuneration		264,000
			Other		21,894
			Total Other Expenses		819,394
		Total Administrative Expenses			819,394
	Total Ordinary Expenses				2,427,847
		Net Assets Variation of the Year			1,170,192
		Net Asset brought forward			7,569,416
		Net Asset carried forward			8,739,608

Publications

[1] Hiroshi Yamakawa, Imagination Architecture and  Consciousness  in the Brain,  in IJCAI-PRICAI 2020 Tutorial T11 ‘Conscious AI: Significance and Development’,  January 8th 2021. https://wba-initiative.org/en/12120/

[2] Mei Sasaki, Naoya Arakawa, Hiroshi Yamakawa, Construction of a whole brain reference architecture (WBRA), International Symposium on Artificial Intelligence and Brain Science, P-31, 10-12, October, 2020. http://www.brain-ai.jp/symposium2020/posters/

[3] Naoya Arakawa & Hiroshi Yamakawa, The Brain Information Flow Format, The 1st Asia-Pacific Computational and Cognitive Neuroscience (AP-CCN) Conference, September 26-27,  2020.
Presentation: https://www.youtube.com/watch?v=0fqdCVCIKC0&t=16s

[4] Yamakawa, H. (2021). The whole brain architecture approach: Accelerating the development of artificial general intelligence by referring to the brain. In arXiv [cs.AI]. arXiv. http://arxiv.org/abs/2103.06123

[5] Taniguchi, T., Yamakawa, H., Nagai, T., Doya, K., Sakagami, M., Suzuki, M., Nakamura, T., & Taniguchi, A. (2021). Whole brain Probabilistic Generative Model toward Realizing Cognitive Architecture for Developmental Robots. In arXiv [cs.AI]. arXiv. http://arxiv.org/abs/2103.08183

[6] Taniguchi, A., Fukawa, A., & Yamakawa, H. (2021). Hippocampal formation-inspired probabilistic generative model. In arXiv [cs.AI]. arXiv. http://arxiv.org/abs/2103.07356

[7] Yamakawa, H. Attentional Reinforcement Learning in the Brain. New Gener. Comput. 2020. doi:10.1007/s00354-019-00081-z

[8] Ayako Fukawa, Takahiro Aizawa, Hiroshi Yamakawa and Ikuko Eguchi Yairi, Identifying Core Regions for Path Integration on Medial Entorhinal Cortex of Hippocampal Formation, MDPI Brain Science, 10(1), 28, 2020.  https://www.mdpi.com/2076-3425/10/1/28

[9] Yoshimasa Tawatsuji, Naoya Arakawa, Hiroshi Yamakawa, Knowledge representation for neural circuits subserving saccadic eye movement based on a Brain Information Flow description, International Symposium on Artificial Intelligence and Brain Science, P-45, 10-12, October, 2020. http://www.brain-ai.jp/symposium2020/posters/

[10] Hiroshi Yamakawa, “Understanding the computational meaning of the neocortical interrarea signals”, CSHL Meeting: From Neuroscience to Artificially Intelligent Systems (NAISys), November 9 – 12, 2020.

[11] Yamakawa, H. Revealing the computational meaning of neocortical interarea signals.  Frontiers in Computational Neuroscience. (2020) https://www.frontiersin.org/articles/10.3389/fncom.2020.00074/full

[12] Hiroshi Yamakawa, Towards a qualitative evaluation of biological plausibility for brain-inspired software, The 1st Asia-Pacific Computational and Cognitive Neuroscience (AP-CCN) Conference, September 26-27,  2020.
Presentation: https://www.youtube.com/watch?v=mtXcFHlKbFw&t=9s

[13] Hiroshi Yamakawa, Naoya Arakawa and Koichi Takahashi, Whole brain reference architecture to evaluate biological plausibility of human-like artificial intelligence,  International Symposium on Artificial Intelligence and Brain Science, P-30, 10-12, October, 2020. http://www.brain-ai.jp/symposium2020/posters/

[14] Heecheol Kim, Masanori Yamada, Kosuke Miyoshi, Tomoharu Iwata, Hiroshi Yamakawa, Reinforcement Learning in Latent Action Sequence Space, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2020), October, 2020. http://ras.papercept.net/images/temp/IROS/files/0338.pdf