February 2023

Revised April 2024

Brain-morphic Software Design Guide

Brain Reference Architecture Design based on the SCID method

The Whole Brain Architecture Initiative

Introduction

Purpose of this document

This guide shows how to create and publish BRA data, or brain-morphic software design data, for Brain Reference Architecture (BRA)-driven development.

* Note: Japanese version of Brain-morphic Software Design Guide can be found here.

What is BRA-driven development?

BRA data can be submitted to the BRAES (BRA Editorial System) site.  BRA data submitted and accepted through the review process are published and shared on the site.

Overall Flow through Data Creation to Posting

Figure. 2 Procedure from data creation to publication

Data to be Posted

Data consists of the following types of data.

* If it is only data related to anatomical structure (BIF data), there is no need to create or post it.

In this guide, we will explain the following points in order:

• How to design BRA data

• Concepts related to BRA data design
• For specific data creation methods, refer to the manual:

• BRA Data Preparation Manual-en.

• How to create BRA image/FRG image
• How to publish BRA data

How to Design BRA Data

BRA Data Structure

The BRA data format consists of the Brain Information Flow (BIF) diagram format, which describes anatomical structures, and the Hypothetical Component Diagram (HCD) format, which describes computational functions consistent with the structure.  The Structure-Constrained Interface Decomposition (SCID) method [Yamakawa 2021] would be used as a method for constructing HCD consistent with BIF, in which a Function Realization Graph (FRG), or a functional hierarchy diagram, is designed.

As shown in Table 1 below, the information on BIF is described in the BIF-related (References, Circuits, Connections) sheets on a spreadsheet and in a BIF Image.  HCD information constructed consistently with the BIF data is described in the FRG sheet and HCD and FRG Images.

Table 1: Types of data describing BIF and HCD

Below, an overview of BIF, HCD, and FRG is provided, and the next chapter will explain the specific design method.

Brain Information Flow (BIF)

Brain Information Flow (BIF) diagram is a directed graph to represent anatomical structures in the brain to various granular levels, in which “Circuits” as nodes are connected with “Connections” (axon projections) as links between nodes (see Figure 2). BIF is described in the BIF-related sheets (References, Circuits, and Connections) on the BRA data and in a BIF Image.

A Circuit is a node associated with a connected neural circuit in the BIF graph structure.  It may represent areas such as the entire visual cortex, the V1 (primary visual cortex), or the neocortex-basal ganglia loop.  The areas covered by Circuits may overlap.  As explained in Section 3, a Uniform Circuit is designed as a group of neurons consisting of a specific cell type. Uniform Circuit is the smallest unit of Circuit, and it is the only starting point of output connection (projection). Thus, all other Circuits include one or more Uniform Circuits.  The attributes to be registered for each Circuit include CircuitID, Source of ID, Names (Circuit aliases), Sub-Circuits, Super Class, Uniform (whether it is a Uniform Circuit).

A Connection corresponds to a link in BIF data, and represents a bundle of axons that transmits signals between Circuits in the brain, as shown in the link in Figure 4A.  Since the starting point of a connection is limited to a Uniform Circuit, the projection destination is described as a list in the Projections attribute only for Circuits for which the Uniform attribute is set to True.  The information described for each projection destination includes the Circuit of the destination, the (average) number of axons, and the feedforward/feedback direction between neocortical regions.

Describing quantitative values ​​in BIF data, such as the approximate number of neurons in a circuit (Size) or the approximate number of axons in a connection (Size), is somewhat complicated, but it brings several advantages. First, it can be used as a reference when implementing computational models where the amount of signals exchanged between components is an issue (in a Leaky Integrated Fire model, or Artificial Neural Network model, for example).  Second, being able to estimate the upper limit of the amount of information to be transmitted often provides hints for considering the computational functions performed by components before and after transmission.

Figure 3 Brain Information Flow (BIF) format

A: Conceptual diagram of BIF. B: An enlarged view of the leftmost circuit in A and a schematic depiction of the two types of neurons within it. C: Attribute representing each circuit in BIF data.

Hypothetical Component Diagram (HCD)

In a hypothetical component diagram (HCD) functions are decomposed to match the structure of the BIF so that the tasks and functions performed by the brain regions of interest (ROI) on the BIF can be accomplished.^[1] In order to match the structure of HCD with that of BIF, HCD is created by reusing the structure of BIF.  Specifically, a component on any HCD is associated with a specific circuit on the BIF, and a dependency on any HCD is associated with a specific connection on the BIF.  HCD is described on the FRG sheet of BRA data and in an HCD Image.

Because HCD is hypothetical in nature, there is little guarantee that it matches the truth about the brain. By allowing multiple HCDs to be assigned to circuits on any BIF, it is possible to describe hypotheses that are inconsistent with each other.  On the other hand, a common HCD has also been established to describe highly reliable functions (for example, coordinate systems).

Figure 4 BRA description example

A, B, C, and D correspond to each other.

Function Realization Graph（FRG）

In order to build software that functions like a brain, we need to move from the parts (what functions are achieved by the smaller parts) to the whole (what functions are achieved by the software parts or the whole). In addition to bottom-up design in which functions are piled up, functions are decomposed from the whole (what functions are achieved by software parts or the whole) into parts (what functions are achieved by smaller parts). Functional hierarchy diagram in parallel with top-down design^[2]  Reverse engineering to create is an effective approach, and the SCID method described below is also a design methodology based on this bi-directional approach. In such bidirectional design, it is necessary to distinguish whether the functional description in the design specification originates from bottom-up or top-down design.  This is because it becomes difficult to know whether the function is present or not, leading to confusion in understanding and evaluating the function.  The Function Realization Graph (FRG) was proposed as a functional hierarchy diagram that allows bidirectional design to be performed while avoiding the problem.

The FRG is a functional hierarchy diagram that hierarchically shows the dependencies between functional nodes, as shown in FIG 4. Each functional node is identified by a Node ID, and dependencies between nodes are defined through Subnodes. The FRG (Figure 4A) is constructed on the HCD, which is designed to match the brain anatomy. In other words, the leaf node of the FRG is a component on the HCD.

Figure 5. FRG (A) and HCD (B).

The HCD is composed of components that correspond to the terminal (leaf) nodes of the FRG.

A list of node attributes related to specific functional nodes is shown in Table 2. The node attributes that each functional node has that are identified by Node ID are Subnodes, Requirement, Capability, Interface, Implementation, Uniform circuits, Output semantics, and Projected circuits.

Table 2. Node list

A list of node attributes related to specific functional nodes is shown in Table 2. The node attributes that each functional node has that are identified by Node ID are Subnodes, Requirement, Capability, Interface, Implementation, Uniform circuits, Output semantics, and Projected circuits.

Taking FIG. 4 as an example, Node ID A and Node ID δ correspond to Subnodes for Node ID β. Requirement is a required function for this Node IDβ to achieve a higher-level function.

Capability is a function realized by its neuroscientific structure, and is distinguished from Requirement. Typical Capabilities include functions realized by Network Motifs (see the Column below).

BRA Data Design: SCID Method

BRA design is performed using the Structure-Constrained Interface Decomposition (SCID) method, which is the core method for BRA-driven development.

The SCID method constructs a BIF for a specific ROI (region of interest) based on anatomical knowledge, and then uses the BRA to achieve the top-level function (TLF) of the ROI to match the BIF.  It is a reverse engineering method to construct BRA [Yamakawa, 2021].  BRA data consistent with neural circuits obtained with the SCID method can be used as design information for brain-morphic software.

Figure 8：Structure-constrained Interface Decomposition (SCID)

BRA designers using this method would follow the following steps to create a functional hierarchy that is consistent with the anatomical structure of the neural circuit of interest (ROI) and achieves the top-level function of the ROI.

• Step 1: BIF construction: Create BIF based on research on neuroscience findings in ROI. 
• Step 2: Determine consistency between ROI and TLF.

• Including provisionary component diagram creation

• Step 3: HCD creation

• Step 3-A: Enumeration of candidate component diagrams (covering possibilities)
• Step 3-B: Reject HCD that is inconsistent based on scientific findings

*In the above, Step.1 and Step.2 are considered in relation to each other.

The steps will be described below.

Step 1: BIF Construction

In this step, we investigate and collect anatomical information such as human or non-human mammalian connectomes around the assumed ROI. The anatomical structure to be focused on as an ROI is described as a BIF Image.  For the notation, see "Creating a BRA image."  Here, we are to create BIF of what is called Supplemented human connectome with other mammals (SHCOM). SHCOM is a hypothesis of a human mesoscopic anatomical structure obtained by integrating current neuroscience knowledge.  As a general rule, we use brain region (Circuit) IDs from the Allen Developing Human Brain Atlas Ontology (DHBA) [Ding, 2017].  If an appropriate element to be mapped cannot be found in DHBA, a new element can be added. do.  For the neocortex, SHCOM is constructed from humans and non-human apes, and for subcortical regions, SHCOM is constructed by mainly referring to rodents.

BIF data for the entire brain has been created as WholeBIF(to be renamed to WholeSHCOM) with DHBA IDs.  In the process of the SCID method, you can add Circuits and Connections if necessary.  For information on how to add them, see "Creation of BIF data" in BRA Data Preparation Manual-en.

Step 2: Consistency Determination of ROI and TLF

In this step, while investigating scientific knowledge on the cognitive behavior of humans and animals, we determine the brain region (ROI) to be designed and the TLF that is consistent with the input/output signals to the ROI.

Generally, if the scope of the ROI is narrow and the BIF network included therein is simple, associated functional hypotheses would be numerous, and the constraints are too weak. In the Step 3, it would be better to secure a reasonably wide ROI in order to effectively reject biologically inappropriate functional hypotheses.

The TLF determined by HCD is written in the Requirement column of the Node ID corresponding to the root node on the FRG sheet of BRA data.

In this step, it would be necessary to guarantee the input/output specifications of the TLF (the ability to convert the Output Semantics of the input signal to the Output Semantics of the output signal) in the assumed TFL in some way.  The following methods are available for this purpose.

• Demonstrate existing operational computational mechanisms and models that meet ROI input/output specifications
• Design and present a provisional component diagram,
  which does not have to match the brain anatomy.

If the ROI determined as the target is not registered as a circuit in WholeBIF, it will be necessary to create a new one on the BRA data Circuit sheet.

Step 3-A: HCD Creation and Description

In this step we will design software components to match the structure of circuits and connections on BIF, and create a mechanism that can achieve TLF through their cooperation.

HCD Image (HCD layer) and BRA data are created for this purpose (see BRA Data Preparation Manual-en).  Even though there is not necessarily a clear hierarchical structure in the anatomical structures on the referenced BIF, it is desirable that the TLF be functionally decomposed hierarchically so that it is easier for HCD users to understand.

While the component decomposition for a TLF is common in software development, the current method differs in that the architecture is constrained by BIF.

Ideas for making output components compatible with Uniform Circuits

In HCD, there is a restriction that the sources of connections (projections) are limited to Uniform Circuits. In order to satisfy this constraint, the component corresponding to the non-Uniform Circuit needs to make output from the components corresponding to the Uniform Circuits inside it.

If Uniform Circuits convenient for this purpose are found inWholeBIF, you can use them.  if such Uniform Circuits are not registered, it is necessary to create new Uniform Circuits on the BRA data circuit sheet.  It is desirable to set the Uniform Circuit so that it is supported by some kind of anatomical knowledge, but if it is difficult, define a temporary Uniform Circuit and use the corresponding component as the output source. The value of the Socuce of ID column for this newly added Uniform Circuit will often be “makeshift”.

Items to be designed as HCD

For HCD design, the following tasks will be performed and will be written on the FRG sheet. Regarding the description method, see BRA Data Preparation Manual-en .

Deciding which components to use and their positioning

Determine the Circuit to be used as a component in the HCD.

The results are recorded by writing the Circuit ID in records in the FRG sheet.

Determining the label of a component

Determine a label that clearly expresses the function of the component.  The label is also used on HCD Image.

The results are recorded by writing the character string of the determined component label in the Component Label (HCD number) column of records on the FRG sheet.

Determining the output destination circuit (Uniform Circuit compatible components)

Only when the component supports Uniform Circuit, select one or more components to be projected from among the circuit identifiers listed in the Circuit ID column on the FRG.

The results are recorded by writing the determined output circuit identifier in the Output Circuits (HCD number) column of records on the FRG sheet.

Determine component calculation mechanism

Determine the process that realizes the transformation from the input from another component (described by I (input circuit identifier)) to the output of this component (described by O (circuit identifier of this component)) .

The results are recorded by writing the determined process in the Process (HCD number) column of records on the FRG sheet.

Determining Output Semantics

Determines the external interpretation of this component's output (described by O (this component's circuit identifier)).

The results are recorded by writing the determined Output Semantics in the Output Semantics (HCD number) column of records on the FRG sheet.

Determination of hierarchical structure (FRG: Function Realization Graph design)

To make it easier for HCD users to understand, it is desirable that TLF be decomposed hierarchically.  Fill in the necessary information on the FRG sheet based on the idea of ​​Function Realization Graph (FRG) was devised for this purpose (👉 Function Realization Graph(FRG)).

Step 3-B: HCD Selection

Here, we reject candidates for HCD that are logically contradictory based on scientific knowledge from various fields such as neuroscience, cognitive psychology, evolution theory, and developmental theory.

It is desirable that a unique HCD be determined as a result of these considerations. However, what should be more important is to compile a comprehensive list of HCD candidates that will avoid rejection. The reason is not only that the knowledge available from sciences is insufficient, but also because it would be beneficial for a collaborative effort to build a whole-brain HCD.  That is, by preparing multiple HCD candidates, it becomes easier to avoid running out of possible candidates for a certain ROI during subsequent ROI expansion or integration with other HCDs (see HCD integration).

In other words, if the HCD is unique because the possible HCDs are not sufficiently covered, it can be considered that the SCID method is not yet complete and Step 3-A is still in progress.

Step 4: BRA Data Completion

Perform a formal review yourself using the BRA Data Review Tool.  Refer to the following document for information on how to use the Review Tool and the results of the formal review.

• BRA Data Review Tool Manual : BRA Data Review Tool Instruction Manual

This format review is automatically performed immediately after data submission (👉 Posting BRA data to BRAES), but we strongly recommend using the Review Tool in advance.

How to make BRA and FRG Images

BRA images consist of a circuit diagram (BIF-image) that expresses the anatomical structure of the ROI and a component diagram (HCD image).  FRG image is a diagram representing FRG. The format is in xml.  Therefore, we recommend creating using draw.io (explanation follows).

Creating a BRA Image

In BRA image, BIF image and HCD image are created separately in layers. You can use any software that can create an xml file, but this guide recommends draw.io.  The explanation will be based on the drawing method using the layer function. In addition, the appendix provides a brief explanation of how to use drawi.io.

BIF and HCD images are drawn in different layers using draw.io. A BIF image is drawn in the BIF layer, and an HCD image in the HCD layer, and a Semi-Transparent layer inserted between these two layers.  When the BIF layer and HCD layer are displayed at the same time, an opaque rectangle on the Semi-Transparent layer is drawn to make it easier to see the contents of the HCD layer and hide the contents of the BIF layer.

When drawing the contents of the layers, follow the instructions below.

BIF Layer

• ROI

• It is shown in rectangular brackets and contains the circuit structure of interest.
• Make it clear that it is ROI.

• Circuit/Uniform Circuits

• Show it as a rectangle and write the Circuit ID in text inside. The position of the text does not matter as long as it is inside, but it should be designed to increase visibility.
• Non-Uniform Circuits must be drawn so that the Uniform Circuit is included.

• Connections

• Write BIF excitatory connections with black arrows.
• Write BIF Inhibitory connections with blue arrows (terminal points are blue squares).
• Write BIF Modulatory connections with green arrows.
• In the case of mutual connections, if the connections are of the same type, a double-headed arrow can be used.
• The thickness of the arrow is as follows.

• 2-point solid arrow: strong projection with anatomically and physiologically recognized connection
• 1 point solid arrow: connection is anatomically/physiologically recognized but weak projection
• 1-point dotted arrow: projection whose connection is hypothesized anatomically and physiologically

Semi-Transparent layer

A Semi-Transparent layer is inserted to draw an opaque rectangle to make it easier to see the contents of the HCD layer when the BIF layer and HCD layer are displayed at the same time, and to cover up the contents of the BIF layer.

Create a rectangle with 60% opacity (no border) and place it so that it covers the content written on the BIF layer.

HCD layer

HCD is drawn on the frontmost layer, which is different from the layer in which the BIF is written, so that the correspondence between the BIF Circuits and Components can be easily understood.

• Component

• Draw an ellipse (or circle).
• Place it over the (Uniform) Circuit in the BIF layer so that it is clear which Component is assigned to which Circuit.
• Inside the oval (or circle), write the label written in the HCD file.

• Connection (connection between components)

• Draw with a white arrow.
• For a two-way connection, use a double-ended arrow symbol.

Example of HCD image

An example of an HCD image created with draw.io is found here:

• HCD Image - Sample.drawio 

Examples of BIF image and HCD image are shown below.

Sample 1：BIF / HCD image

Sample 2：

For other information on how to draw BIF of the cerebral neocortex in HCD image and examples related to Connection, see HCD Image - Sample.drawio.

Creation of FRG image

A diagram that represents the functional hierarchy of an FRG is called an FRG image. This diagram is created according to the following:

• Node

• Represented by a rounded rectangle
• Write the Node ID in a rounded rectangle

• Link

• Draw an arrow from Node to Subnodes
• The point of the arrow should be 2pt.

Figure. FRG image example

Image Output

Output the created BRA image as the following two xml files.

• BIF image: Output of the content written in the BIF layer as an xml file
• HCD image: Output of the contents written in the HCD layer as an xml file

Image file naming conventions

File names should follow the naming rules below.

• Project name + BIF/HCD/FRG.xml
• Example: YT24OculomotorBIF.xml

Appendix: How to use drawi.io

The following is an overview of how to create BIF/HCD/FRG images with draw.io.

1. draw.io access. When the screen below appears, select "Create a new file".

* If you see a screen like the one below, select "Google Drive" and when it says "Authentication required" select "Authenticate". Once authentication is complete, you will be redirected to the screen shown above.

2. Select the default “Basic/Blank File” and select “Create”.

3. Create the BIF neural circuit structure and HCD component structure in separate layers. Create a layer from "View ⇒ Layer".

4. When the layer window appears, select "Add layer" and add the "BIF" and "HCD" layers. (Double-click the added layer name to change the layer name)

5. Create a diagram. For specific instructions on how to use draw.io, refer to the official page: https://drawio-app.com/tutorials/

6. Select the diagram you created and output it as an xml file.

"File" → "Export as format" → "XML..."

Select "Selected range only" and execute "Export".

For the filename check Image file naming conventions.

Publication of BRA Data

A complete set of BRA data can be posted using the BRA data submission form on BRAES, the BRA portal site,  run by the  Whole Brain Architecture Initiative (WBAI).  Submitted data will be reviewed and the accepted data will be posted on the site under the Creative Common License (CC-BY-SA: Attribution-Share-Alike).

The flow from data submission to publication is summarized below (dotted line in the figure below).

1. Post the set of submission data (*) from the portal site (☞Posting BRA data to BRAES).
2. Two types of reviews and response to correction requests (☞Responding to correction requests from BRAES）

1. Corrective responses to format examinations
2. Request for modification of expert review

figure. Flow from data submission to publication

Posting BRA Data to BRAES

Submit the data you have prepared using the form below.

*For data to be posted, see Data to be Posted.

The contents to be filled in the form are as follows.

Please fill out all the contents in English.

Table. Questions and answers in the submission form

Corrective Responses to Format Review

Immediately after posting, an email message will be sent to the registered email address with the subject line below to inform you that the submission has been accepted.

• Subject: [BRAES ID:{Submission ID}] Thank you for your BRA Data submission

After a while, the format review results will be announced.on mailwill be replied.

• Subject: [BRAES ID:{Submission ID}] Your data passed the format review

• After passing the format review, the application will proceed to manual review by experts. You will be notified on the results of this review via email.

• Subject: [BRAES ID:{Submission ID}] Your data failed to pass the format review

• Error was detected during the format review.  Please address the error by referring to the following document.

• BRA Data Review Tool Manual-en (BRA Data Review Tool Instructions)
• BRA Data Preparation Manual-en

* Format review may take some time.  If you do not receive an email after a while, contact  bra-support@wba-support.org.

Responding to CorrectionRequests for Manual-Review

As mentioned, data that passes the format review will be reviewed by experts.  The result
{Accepted/ Conditionally Accepted/ Reject} will be notified via email with the subject below.

• Subject: [BRAES ID:{Submission ID}] We are appreciated to inform the review result

Accepted: BRA data will be published immediately on BRAES (☞ Publication of BRA Data).

Conditionally Accepted: Make corrections according to the expert review results and repost the corrected data from BRAES (☞ Posting BRA data to BRAES).

Reject: We still recommend that you make corrections according to the expert review results and resubmit the corrected data from BRAES.

Manual-review is examined for biological plausibility from the following perspectives.

As shown in Table 2, it consists of three aspects: authenticity of BIF, consistency of HCD with BIF, and functionality of HCD. Authenticity evaluation is performed on the Reference sheet, Circuits sheet, and Connections sheet in the BRA data. Consistency evaluation and Functionality evaluation are performed on the FRG sheet.

Table 2: Evaluation perspective of BRA data

How to Check the Examination Results

The manual review columns that the reviewer writes in Manual-review are the following columns in the four sheets to be reviewed. Check items and make corrections as appropriate.

The explanation of error codes are found in BRA Data Review Tool Manual-en .

Table: Manual-Review Columns: Columns written by Reviewer in each sheet

Publication of BRA Data

BRA data accepted in the manual-review will be published under the CC-BY license on the BRAES site.

   BRA Data Official Portal Site (BRAES):
          https://sites.google.com/wba-initiative.org/braes/

References

Read the following paper to learn BRA-driven development as appropriate.

[Yamakawa 2021] Yamakawa, H. (2021). The whole brain architecture approach: Accelerating the development of artificial general intelligence by referring to the brain. Neural Netw. 144, 478–495. https://doi.org/10.1016/j.neunet.2021.09.004

[Yamakawa 2022] Hiroshi Yamakawa (2022). Whole-brain architecture ─ Designing and developing brain-like AI while understanding its functions ─. In Cognitive Science Course 4 Developing a framework that captures the mind Cognitive Science Course., Kazuhiko Yokozawa, ed. (University of Tokyo Press), pp. 209–249. https://www.utp.or.jp/book/b609203.html