Neural Morphology AI: The Next Step in Future Artificial Intelligence

2019-10-18

When Apple CEO Tim Cook introduced the iPhone X, he claimed that he would pave the way for the next decade of technology. Although it is still too early, the emergence of neural engines for facial recognition has already proven some issues. Today, deep neural networks have become a reality, and neuro-morphology seems to be the only viable path for advancing AI.


According to a recent report by Yole Développement, sensing and computing must reshape themselves by mimicking neurobiological architectures in the face of data bandwidth limitations and growing computational requirements.


Yole imaging analyst Pierre Cambou said in an interview, "Neuro-morphic sensing and computing can address most of the current challenges in AI and open up new application prospects in the coming decades. Neuro-morphic engineering is the next step in biomimetics and will drive the development of AI."It has been 70 years since mathematician Alan Turing posed the question "Can machines think?"It has been 30 years since Carver Mead, an electrical engineer at the California Institute of Technology, introduced the concept of neuro-morphic engineering.However, in the next decade, researchers have had little practical success in building machines with brain-like learning and adaptability.


The turning point came with the publication of the paper "ImageNet Classification with Deep Convolutional Neural Networks" by a group of scientists from the University of Toronto. The AlexNet architecture consists of 8 layers of convolutional neural networks and can classify 1.2 million high-resolution images from the ImageNet competition into 1,000 categories (e.g. cats, dogs). It was only with the development of AlexNet that deep learning methods were proven to be more powerful and began to gain momentum in the field of AI.Most current implementations of deep learning techniques rely on Moore's Law and work well. However, as deep learning advances, the demand for chips that can perform high-performance computing tasks will continue to grow. Moore's Law has recently been slowing down and has led many industry insiders, including Yole Développement, to believe that Moore's Law will not be able to sustain the progress of deep learning. Cambou believes that deep learning will ultimately fail if it continues to be implemented in the same way.


To explain his point of view, Cambou listed three major obstacles. The first is the economics of Moore's Law. Few companies can participate in the competition. He believes that it is harmful to innovation that only Google can do something. Secondly, the growth rate of data loads is faster than Moore's Law, and data overflow makes current storage technologies a limiting factor. Thirdly, the exponential growth in computational power requirements creates a thermal wall for each application. For example, using a 7nm chip, we have an efficiency of about 1 teraflop per watt (1 trillion floating-point operations per second). When powering an electric car, we may need one kilowatt, which means we need one hundred billion terabytes. Current technology cannot achieve this, and the solution may be to apply deep learning on neuro-morphic hardware and improve energy utilization efficiency.Cambou said, "From a broader perspective, the future approach will be a disruptive one that leverages the benefits brought by emerging memory technologies and improves data bandwidth and energy efficiency. That is the neuro-morphic approach. The story of artificial intelligence will continue to evolve, and we believe the next step is towards neuro-morphism."


In recent years, many efforts have been made to build neuro-morphic hardware that conveys cognitive abilities through silicon-based neurons. For Cambou, this is a feasible approach that improves efficiency. We believe it will take neuro-morphic AI to the next level. Then, we can redefine AI.

RECENT AI NEWS

RECENT AI TOOLS