China
Africa
Explore chapters and articles related to this topic
Nanotechnology in Preventive and Emergency Healthcare
Published in Bhaskar Mazumder, Subhabrata Ray, Paulami Pal, Yashwant Pathak, Nanotechnology, 2019
Nilutpal Sharma Bora, Bhaskar Mazumder, Manash Pratim Pathak, Kumud Joshi, Pronobesh Chattopadhyay
The most complex entity in the universe is undoubtedly the human brain and, up to this point in time, no functioning theory of the brain has been universally accepted. In order to replicate the human brain, understanding it is the key to success, but due to a lack of proper technology and machinery, a complete human brain map is still under development. Non-invasive and invasive are the only two methods by which brain mapping may be accomplished. The problem with the non-invasive method is that it does not allow scientists to carry out proper mapping of the human brain in live condition. Invasive mapping technology is built around two techniques, brain implants and nanorobots. Neurosurgery is required to place the brain implant, which stimulates and records the function of neurons via telemetry signals, but the safety profile of the whole process is low in both the long or short term. In the year 2014, IBM announced the most advanced brain-like chip to be developed to date called TrueNorth. The efficiency of this brain-like chip is 768 times more than any other contemporary chip ever built (Flores, 2016; Salinas, 2015).
Exponential Improvement
Published in Chace Calum, Artificial Intelligence and the Two Singularities, 2018
Carver Mead is a pioneer of the microelectronics industry based at CalTech; he gave Moore’s law its name. In the late 1980s, he proposed the idea of computers based on the architecture of the brain, called neuromorphic computing. As with many developments in computing, it could not be realised at the time, but is now becoming a reality. In 2014, IBM produced a chip called TrueNorth, which comprised around a million silicon ‘neurons’, each with 256 ‘synapses’. As well as being very powerful, IBM claimed the chip is extremely energy-efficient.
A Cross-layer based mapping for spiking neural network onto network on chip
Published in International Journal of Parallel, Emergent and Distributed Systems, 2018
To achieve high performance of neural network applications, efficient neural mapping algorithm is required to place neurons into specific NoC cores. In addition, the relative position of different neural layers mapping to on-chip network also should be taken into considered to improve inter-neuron communication efficiency. Current neuromorphic systems are prone to modeling massive neurons and the scale of NoC-based SNN infrastructures becomes huge greatly. The SpiNNaker engine consists of up to 1,036,800 ARM9 cores and each capable of simulating up to 1K spiking neurons [3,22]. The TrueNorth chip presents 4096 parallel cores, where each core realises 256 spiking neurons and 64K synapses [9,10]. As the increasing size of NoC-based SNNs, the complex interconnection among spiking neurons and the number of realised neurons per core, energy consumption and spike transfer latency of neuromorphic systems depends largely on the neural mapping efficiency. Therefore, it is essential to propose a mapping strategy aiming at improving NoC-based SNN system performance. However, only few works related to neural mapping were found in literature. In TrueNorth neuromorphic system [10], total weighted wire-length among neurons are regarded as the main concern of mapping. SpiNNaker [23] uses a sequential mapping method which integrates continuous neurons into one population. There is no connection inside population and the neurons are mapped into on-chip network nodes sequentially. Based on the neural population, Yu et al. [24] propose a swap strategy which makes populations fragmented and swaps neurons with each other in populations with various active degrees so as to obtain a balanced distribution. However, it is difficult to balance the fragmentation and neural active degrees. Mand et al. [25] consider mini-column as basic functional element which consists of 100 neurons and is mapped to each core. Done et al. [26] propose a mapping strategy for NoC-based neural network, which maps neurons layer by layer but different layers’ neurons cannot be placed into the same node. All of the above existing neural mapping strategies follows sequential mapping scheme that maps neural populations into on-chip network cores one by one. The methods deal with inter-neuron communication between different cores poorly.