Intel chip has "smell"! 3000 times better accuracy

Researchers from Intel Research and Cornell University have jointly published a paper in the journal Nature Machine Intelligence, demonstrating Intel's neural mimicry research chip in the presence of significant noise and obscuration "Loihi" Ability to learn and identify hazardous chemicals.

According to reports, Loihi only needs a single sample to learn to recognize each odor without destroying its memory of the odors it has previously learned, and it has shown excellent recognition accuracy.

If you use the traditional method, even the best deep learning solution, to achieve the same odor classification accuracy as Loihi, learning each odor requires more than 3000 times the training sample.

According to Nabil Imam, a senior research scientist in the Neuromimic Computing Group at Intel Research Institute and a PhD in neuromimic computing, Cornell University is responsible for studying the animal's biological olfactory system and measuring brain wave activity when the animals smell the odor. These circuit diagrams and electrical pulses are used to derive a set of algorithms and configure them on a neural mimicry chip.

Imam led the team using a data set consisting of 72 chemical sensor activities to respond to 10 gaseous substances (odors) circulating in a wind tunnel experiment. The sensor's response to various odors is transmitted to Loihi, whose chip circuits simulate the brain circuits behind the sense of smell.

Under the guidance of researchers, Loihi has quickly mastered the neural characterization of 10 different odors, including acetone, ammonia and methane, and can accurately identify even strong environmental disturbances.

Traditional smoke and carbon monoxide detectors can use sensors to detect odors. Although they can detect harmful molecules in the air and issue an alarm, they cannot distinguish between various odors.

"We are developing a neural algorithm on Loihi to simulate how the human brain smells when it smells," said Nabil Imam, a senior research scientist at Intel's Neural Mimic Computing Laboratory. "This work is contemporary research in the intersection of neuroscience and artificial intelligence technology. Exemplary and confirm that Loihi has the potential to provide important perception capabilities and benefit all walks of life. "

AbilNabil Imam holds a Loihi neuromimetic test chip in the neuromimetic computing laboratory in Santa Clara, California

The Intel Loihi neural mimicry chip was born in September 2017. It deviates from the traditional von Neumann computing model of the silicon chip, but instead simulates the neural mimicry of the human brain. It is an asynchronous circuit that does not require a global clock signal. Instead, it uses Asynchronous impulsive neural networks (SNNs) are up to 1000 times faster and up to 10,000 times more energy efficient than traditional CPUs in certain applications.

In July 2019, Intel also announced a new neural mimicry system codenamed "Pohoiki Beach", which contains up to 64 Loihi chips, integrated 132 billion transistors, a total area of 3840 square millimeters, 8 million neurons, 8 billion Synapses.

In theory, Loihi can be expanded to a maximum of 16,384 chip interconnects, which is more than 2 billion neurons-the human brain has about 86 billion neurons.