Brain Intelligence: Go Beyond Artificial Intelligence

In plain words

Today's AI systems are good at one narrow job, such as recognizing images, understanding speech or holding a conversation. They usually need a lot of data, and they do not combine their skills the way a whole brain does. The authors name four gaps: a system is limited to one frame of problems, it does not link patterns to a broader understanding, it struggles to tie symbols to their meanings, and it leaves open how a mind relates to a body.

Their proposal is called Brain Intelligence. Instead of one large network, it is a group of small networks that are wired together the way living cells in a colony would be. Each small network can pass what it has learned to its neighbors, and the group also gets a memory and a way to imagine outcomes of events it has never experienced. The paper describes this as a direction for research. It does not show a working system.

TL;DR

A 2017 position paper by Lu et al. that proposes Brain Intelligence (BI), a network of small neural sub-networks connected by artificial-life links, with memory and an imagine function. The paper specifies no equations, algorithms, hyperparameters or experiments, and it reports no results. Use it as a map of the authors' argument, not as a method to implement.

Problem

The authors argue that current AI is specialized: systems cover one area such as image recognition, speech recognition or dialogue, and do not match whole-brain functions such as self-understanding and self-motivation §1. They list four problems §1:

• Frame problem. Big-data training takes long, so a system is confined to one frame or problem type. Covering every real-world event would mean anticipating an unbounded number of possibilities §1.
• Association function problem. Learned models extract particular patterns and produce numerical outputs, but lack the association of the whole brain; a single part of the brain is not as capable as the whole §1.
• Symbol grounding problem. Linking a symbol to its meaning is often unsolved. The example given is that a zebra is a horse with stripes §1.
• Mental and physical problem. How a non-material mind affects a physical body is not settled §1.

The abstract adds that recent systems depend heavily on big data, lack a self-idea function and are complicated [abstract]. None of these claims is backed by a measurement in the paper. The assumptions are argument and citation: that single-area systems cannot be extended to whole-brain function §1, and that a useful model needs only a small database §3.

Method

The BI model is proposed in §3 as a network built from many simple sub-networks §3. The authors say it combines artificial life (AL) with AI and adds a memory function and an idea function §3 fig. 3. They present it as a concept, and the paper gives no equations or algorithm.

1. Decompose. The BI network is made of many simple sub-networks, each a neural network §3 fig. 3. Figure 3 draws six networks: one top network, two middle ones and three at the bottom fig. 3.
2. Link. Sub-networks are connected by an AL-based network that lets parents and sons share their parameters, trained results and structures fig. 3, caption.
3. Update. Each sub-network's parameters are updated by an S-system. The S-system modifies sub-networks through reproduction, selection and mutation §3. The S-system is not defined in this paper.
4. Memory and imagination. The model includes a memory and an idea (imagine) function §3 [abstract]. No mechanism for either is given.
5. Training rule. The authors plan to combine regional AI methods under a particular rule, especially unsupervised learning, using a small database §3. The rule is not stated.

The paper contrasts its design with NEAT, HyperNEAT, gene regulatory networks, evolving reaction networks and cognitive computing. It argues that these do not use all of the brain's functions and that the S-system improves the structure of current AI models rather than only optimizing parameters §3.

Pseudocode sketch of the description above. The paper contains no algorithm alg.; every step beyond the text is marked as not stated.

# Sketch of the BI network as described in prose  [§3] [fig. 3]
# Not an algorithm from the paper: the text leaves every step open.
# Input:  a set of sub-networks, each a neural network
# Output: the updated sub-networks

def bi_network_update(sub_networks, links):
    # Connect parents and sons through AL-based links  [fig. 3]
    for parent, child in links:
        # Share parameters, trained results and structure  [fig. 3 caption]
        share(parent, child)
    # Update each sub-network with the S-system rule  [§3]
    for network in sub_networks:
        # Reproduction, selection and mutation; operators not stated  [§3]
        network = s_system_update(network)
    return sub_networks

Key equations: none. The paper has no numbered equations §1–§4.

Hyperparameters

The paper gives no hyperparameters. Each value an implementer would need is listed as not stated.

Results

Author-reported. Setup: none. The paper reports no experiments, datasets, models, hardware or baselines §2–§4. Demonstrations on automatic driving, precision medical care and industrial robots are announced as future work [abstract] §3.

Paper vs code

No official code. The paper links no code repository or project page. Its references include a few websites (for example Siri, AlphaGo and the Japanese AI Expo), but none is a code release §references. A GitHub search for "brain intelligence beyond artificial intelligence" returned four repositories (anas-zafar/agi-cognitive-foundations, Kuberwastaken/BrainAI, Aryia-Behroziuan/Other-sources and pedrohiglesias/oliver). None has an owner matching an author of this paper, so none is used.

Pitfalls

• Nothing is implementable as written. The paper fixes no architecture, no update equations and no training procedure. Sub-network count, link rules, S-system operators, the memory design and the imagine mechanism are all not stated §3. An agent building from this card has to invent them, so treat any implementation as the implementer's own design.
• The S-system is undefined here. The paper uses "S-system" as the update mechanism and cites only reference [62], a 1998 paper on an artificial-life "bug type" §3 §references. Do not assume a formalism from the name alone.
• Figure 3 is ambiguous. Its arrows are bidirectional between the levels and each network carries a self-loop. The caption does not explain the loops or the direction of sharing fig. 3.
• Assertions without evidence. The conclusion says it is foreseeable that BI solves the four problems §4. It also says the S-system has the benefits of an association function §4, and that HyperNEAT cannot overcome NEAT's drawbacks §3. None of these is supported by an experiment or a derivation in the paper.
• Terminology. The paper expands "RL" as "Represent Learning" in §1 and in Figure 1 ("Represent Learning" label). In the field RL usually means reinforcement learning, so do not map the paper's RL to that without checking the text.
• Citation mismatches. AlexNet is cited as [42], but the reference list gives [42] as Karpathy et al., a video-classification paper, and AlexNet is not in the list §2.9 §references. "Dependency neural network (DCNN) [53]" is described as a convolutional layer on top of an LSTM, while [53] is Chen and Manning's dependency parser §2.8 §references. Verify any citation before relying on it.
• Costs are absent. No compute, data volume, memory or training time is given for any part of BI §3.
• Figures. Figure 2's caption credits its content to [31], so it is not the authors' own work; no figure from this paper is reused here, and all figures are described in words only.

Lineage

• Builds on: none stated. The paper's background is the 2017 state of deep learning and related AI areas it surveys in §2 (speech, visual recognition, NLP, hardware, robotic process automation, decision management).
• Compared against: NEAT [63, 64], HyperNEAT [65], gene regulatory networks [66], evolving reaction networks [67] and cognitive computing [68]. The S-system [62] is named as the update mechanism rather than compared.
• Superseded by: not stated.