Source-linked AI summary
Brain Intelligence: Go Beyond Artificial Intelligence
Huimin Lu, Yujie Li, Min Chen, Hyoungseop Kim, Seiichi Serikawa
TL;DR
The paper addresses limitations of recent AI, including dependence on big-data training and restriction to a single frame or problem type. It proposes Brain Intelligence, a model fusing artificial intelligence and artificial life, and concludes that BI can solve several identified AI issues.
Problem
Recent AI is limited by dependence on big-data training and restriction to a single frame or type of problem.
Method
The paper proposes the brain intelligence model, which fuses artificial intelligence and artificial life.
Results
The paper concludes that the BI model can solve issues including the frame problem, association function problem, symbol problem, and self-motivation.
Takeaways & Limitations
The paper positions BI as a model intended to direct future work on the problems of recent AI.
Takeaways & Limitations
Recent AI is typically limited to a single frame or type of problem because of its dependence on big-data training.
Abstract
from arXiv · showhide
Artificial intelligence (AI) is an important technology that supports daily social life and economic activities. It contributes greatly to the sustainable growth of Japan's economy and solves various social problems. In recent years, AI has attracted attention as a key for growth in developed countries such as Europe and the United States and developing countries such as China and India. The attention has been focused mainly on developing new artificial intelligence information communication technology (ICT) and robot technology (RT). Although recently developed AI technology certainly excels in extracting certain patterns, there are many limitations. Most ICT models are overly dependent on big data, lack a self-idea function, and are complicated. In this paper, rather than merely developing next-generation artificial intelligence technology, we aim to develop a new concept of general-purpose intelligence cognition technology called Beyond AI. Specifically, we plan to develop an intelligent learning model called Brain Intelligence (BI) that generates new ideas about events without having experienced them by using artificial life with an imagine function. We will also conduct demonstrations of the developed BI intelligence learning model on automatic driving, precision medical care, and industrial robots.
1. Introduction
Current AI has advanced rapidly and supports specialized applications, but remains limited by narrow task scope, dependence on large-scale data, and unresolved whole-brain functions. The paper reviews these limitations and introduces Brain Intelligence as a proposed next-generation architecture.
- AI progress and applications: Current AI has developed rapidly and is being applied to areas including language processing, visual recognition, autonomous systems, and robotics.Examples include speech and image recognition, autonomous delivery robots, facial recognition, robot cars, cleaning robots, and walking robots.
- AI progress and applications: Weak AI is designed for special tasks and may outperform humans in specific domains, whereas general AI would address nearly every cognitive task.Examples of specialized tasks include facial recognition, internet searches, driving, chess, and equation solving.
- Limitations of Artificial Intelligence: Current AI remains limited to specific intellectual areas such as image recognition, speech recognition, and dialogue response.CNNs and ResNet support visual recognition, while RNNs, DNNs, and representation learning support speech or dialogue-related functions.
- Limitations of Artificial Intelligence: AI has not reproduced whole-brain functions such as self-understanding, self-control, self-consciousness, and self-motivation.The paper also identifies unresolved questions concerning relationships between mind and body, and concludes that many AI problems remain unsolved.
- Limitations of Artificial Intelligence: Big-data training typically confines AI to a single frame or problem because anticipating every real-world possibility would overload the database.The paper states that real-world phenomena contain infinitely many possibilities, making exhaustive extraction time effectively infinite.
- Limitations of Artificial Intelligence: Large-scale data and numerical processing help AI extract particular patterns, but current systems lack the human brain’s association function for connecting ideas.The paper illustrates this limitation with the inability to infer that a zebra is a horse with stripes from separate word meanings.
- Brain Intelligence: The paper reviews weak-AI algorithms and introduces Brain Intelligence as an advanced architecture intended to address disadvantages of weak AI.The proposed architecture is presented as a next-generation intelligence model for solving limitations of weak-AI algorithms.
2. Artificial Intelligence
Artificial intelligence encompasses established and emerging technologies used across applications, but current methods face challenges balancing data-driven capability with efficient, robust output. The section surveys AI techniques including deep learning, natural language generation, and related architectures.
- AI adoption and investment were projected to increase substantially, with enterprise use rising from 38% in 2016 to 62% in 2018 and the market reaching $47 billion in 2020.The cited projections describe rapid growth in enterprise adoption, investment, and market size.
- AI technologies range from established tools to newer methods and are applied across diverse artificial-intelligence tasks.The section frames AI as a broad and changing technology field.
- Deep neural networks learn increasingly abstract representations through architectures and training methods including backpropagation, feedforward networks, log-bilinear models, and recurrent neural networks.These models can represent voice sequences of varying lengths, linking similar histories through related representations.
- Long short-term memory architectures extend recurrent neural networks and support applications such as customer service, report generation, and business-intelligence summarization.LSTM-based encoder-decoder architectures have been proposed for content selection and realization.
- Natural language generation must balance adequate textual output with efficient and robust text generation.The section identifies this balance as difficult in many existing methods.
2.2 Speech Recognition
Speech-recognition research uses statistical and neural architectures to convert human language into formats useful for computer applications. Deep recurrent models, including joint acoustic and linguistic modeling, are widely used in interactive voice systems and mobile applications.
- Hidden Markov models and recurrent neural networks are used for speech recognition, although the cited HMM-RNN model does not perform as well as deep networks.
- Speech recognition aims to translate human language into a useful format for computer applications.
- Graves et al. proposed a deep long short-term memory RNN method that jointly trains separate acoustic and linguistic RNNs.
- The deep LSTM approach is widely used in interactive voice-response systems and mobile applications.
- Virtual and augmented reality systems simulate interactive three-dimensional environments and are associated with future AI applications in remote eHealth.
2.4 AI-optimized Hardware
AI workloads increasingly rely on large datasets and substantial computing power, creating demand for hardware acceleration beyond current data and model sizes. GPUs, GPGPUs, and FPGAs provide different forms of computational support, with FPGAs offering flexible configuration but greater programming difficulty.
- AI models require large amounts of data and computing power to train, motivating hardware acceleration for scaling beyond current data and model sizes.
- GPUs, GPGPUs, and FPGAs are used to run AI-oriented computational tasks efficiently.
- GPUs provide many more computational cores than traditional general-purpose processors and enable greater parallel computation.
- FPGAs offer flexible hardware configuration and better performance per watt than GPUs, but their special architecture makes them difficult to program.
2.5 Decision Management
Decision-making is important for sustainable development in turbulent financial markets, and improved ICT has enabled AI-based decision-making techniques. Decision-management engines combine rules and logic with AI systems for setup, training, maintenance, and automated enterprise decisions.
- Decision-making plays a critical role in sustainable development during turbulent financial markets.
- AI-based decision-making techniques include decision trees, support vector machines, neural networks, and deep learning.
- Decision-management engines insert rules and logic into AI systems during initial setup, training, maintenance, and tuning.
- The technology is mature and is used across enterprise applications to assist with or perform automated decision making.
2.6 Deep Learning Platforms
Deep learning platforms rely on large datasets and substantial computational infrastructure, while current systems remain less efficient than the human brain in perception and resource use. The section surveys hardware and software platforms developed to train deep neural networks.
- Large-scale deep-learning platforms commonly process big datasets on commodity CPU clusters, while GPUs are prominent training platforms.
- Current AI platforms remain worse than the human brain in perception and require substantial space and energy.
- Google’s DistBelief trains deep neural networks using thousands of CPUs, whereas Microsoft’s Project Adam targets training with fewer machines.
- Related hardware includes Qualcomm’s Zeroth, IBM’s TrueNorth, Manchester University’s SpiNNaker, and Google’s TPU.
- Deep-learning software packages include TensorFlow, Theano, Torch/PyTorch, MxNet, Caffe, and the higher-level package Keras.
2.7 Robotic Process Automation
Robotic process automation uses software and algorithms to automate human actions across applications, especially where manual execution is costly or inefficient. Its applications include finance, treasury trading, and data analysis.
- RPA uses software robots to automate typing, clicking, and data analysis across different applications.
- RPA is used for tasks or processes that are too expensive or inefficient for humans to execute.
- Researchers are promoting RPA adoption in finance, including treasury trading that affects accounting staff in banking.
- AI is presented as a solution for big data that provides a possibility for accurate RPA prediction.
- Sample RPA vendors include Advanced Systems Concepts, Automation Anywhere, Blue Prism, UiPath, and WorkFusion.
2.8 Text Analytics and NLP
Text analytics and NLP apply machine-learning methods to help computers understand, analyze, and derive meaning from human language. The section reviews recurrent, convolutional, recursive, and dependency-based neural architectures for language tasks.
- NLP facilitates text analytics by helping computers understand, analyze, and derive meaning from human language.
- RNNs use sequential information by repeating the same task across sequence elements while making outputs depend on previous computations.
- Bidirectional RNNs incorporate future as well as previous sequence elements, and deep bidirectional RNNs add multiple layers per time step.
- LSTM uses a mechanism to decide what information to keep or erase from memory in recurrent neural networks.
- Reviewed alternatives include recursive networks for structured prediction, DCNNs for long-distance dependencies, and dynamic k-max pooling networks for selecting maximum sequence values.
- Other reviewed NLP architectures include multi-column, ranking, and context-dependent CNNs, with the latter combining sentence summarization and representation matching.
2.9 Visual Recognition
Visual recognition methods use deep neural networks, especially CNNs, to learn image relationships and support location-invariant recognition. The section reviews CNN architectures and their trade-offs in depth, computational complexity, and spatial information.
- Deep learning is presented as one of the best solutions for computer vision, using many parameters and local connectivity to learn relationships between neighboring pixels.
- CNN pipelines use convolutional, pooling, and fully connected layers to generate feature maps, reduce dimensions, and produce one-dimensional feature vectors.
- CNNs use learnable weights and biases, with multiple convolutional layers and nonlinear activation functions distinguishing them from traditional neural networks.
- AlexNet contains eight layers and combines convolutional and fully connected layers with data augmentation, dropout, ReLU, normalization, and overlapping pooling.
- VGGNet increases depth through small convolution filters, reaching 16–19 layers, but max-pooling layers lose accurate spatial information.
- GoogLeNet increases network width and depth within a constant computational budget and can be 2–3 times faster than similarly performing networks, but its design is complex to configure.
- As network depth increases, training accuracy can saturate and rapidly degrade; residual learning addresses this with shortcut connections that add neither parameters nor computational complexity.
3. Brain Intelligence (BI)
The BI model is proposed as a general-purpose intelligence framework that combines artificial intelligence with artificial life, memory, concept understanding, and whole-brain functions. Its network uses multiple sub-networks whose parameters and structures can evolve through reproduction, selection, mutation, and S-system-based modification.
- Concept and motivation: The paper proposes Brain Intelligence as a model combining artificial intelligence and artificial life to extend current AI capabilities.The model is intended to address limitations of AI systems focused on individual areas and pattern extraction.
- Concept and motivation: BI targets learning with a small database, concept understanding, memory, and whole-brain functions.The authors identify insufficient research on perceptual understanding and self-thinking models.
- Network design: The BI model network combines artificial life technology, artificial intelligence technology, an idea function, and memory function.Different neural networks are connected within the proposed BI network.
- Network design: BI consists of many simple sub-networks whose parameters are updated by an S-system through reproduction, selection, and mutation.The proposed network is described as an evolving structure rather than only a fixed neural architecture.
- Relation to existing models: Unlike NEAT, the proposed BI model uses S-system-based structure improvement in addition to neural-network structure and parameter optimization.The paper positions this as a distinction from NEAT and related artificial-life-based neural networks.
- Future direction: The paper frames BI as an engineering investigation toward a super-intelligent brain function model that can discover problems autonomously.This future model is intended to incorporate functions beyond current individual-area AI systems.
4. Conclusion
The conclusion reviews current AI tools and problems, then presents BI as a model that fuses artificial intelligence with artificial life. It states that BI may address several identified intelligence problems, including frame, association-function, and symbol-grounding problems.
- Conclusion: The paper reviews state-of-the-art AI tools for individual application areas, including natural language processing and visual recognition.It presents the review as an overview of current deep learning methods and their applications.
- Conclusion: The paper consolidates problems in recent AI models and proposes the brain intelligence model as a direction for future research.The conclusion states that this synthesis can direct future work for researchers.
- BI contribution: The brain intelligence model fuses artificial intelligence and artificial life.Artificial-life models such as the S-system are described as providing an association function distinct from GAN-based big-data generation.
- BI contribution: The paper states that the BI model can solve issues involving the frame problem, association function, and symbol grounding.This is presented as a foreseeable capability of the proposed model rather than a reported quantitative evaluation.