Intel announces new progress in neural mimicry: surpassing traditional CPUs by 1000 times

Recently, Intel shared new developments in the Intel Neuromimetic Research Community (INRC), with Lenovo, Logitech, Mercedes Benz, and machine vision sensor company Prophesee joining in to explore the value of neural mimetic computing in business use cases.

The INRC community was founded in March 2018 and now has over 100 members. Its original intention was to collaborate with the industry to effectively unleash the full potential of neuromimetic computing. In the coming years, we will develop it from a research prototype to a product that can be industrialized.

Previously, in September 2017, as a research project of Intel Research Institute, Intel released the first autonomous learning neural mimicry chip codenamed "Loihi", which includes 128 small cores, each with 1000 neuron hardware, simulating multiple "logical neurons", and improving energy efficiency by 1000 times compared to AI trained general-purpose chips.

In July 2019, Intel released a neural mimicry system codenamed "Pohoiki Beach", which includes 64 Loihi research chips and has 8 million neurons.

In March 2020, Intel demonstrated Loihi's ability to learn and identify hazardous chemicals in the presence of significant noise and cover, requiring only a single sample to learn to recognize each odor.

At the same time, Intel announced a data center rack mounted system codenamed "Pohoiki Springs", which integrates 768 Loihi chips in 5 standard server sized chassis and has 100 million neurons, approximately the size of a small mammalian brain.

This time, the benchmark update for Loihi neuromimicry research, which Intel Day focuses on, includes:

-Voice command recognition

Accenture tested the ability to recognize voice commands on Intel Loihi and standard GPUs and found that Loihi not only achieved accuracy similar to GPUs, but also improved energy efficiency by more than 1000 times and response speed by 200 milliseconds.

Mercedes Benz is exploring how to apply these results to reality, such as adding new voice interaction commands to cars.

-Gesture recognition

Accenture demonstrated the tangible progress Loihi has made in rapid learning and personalized gesture recognition. With just a few exposures, Loihi can learn new gestures that can be used for intelligent product interaction or non-contact display in public places.

-Image retrieval

Researchers in the retail industry evaluated Loihi's application in image-based product search and found that while maintaining the same level of accuracy, Loihi's efficiency in generating image feature vectors is more than three times higher than traditional CPU and GPU solutions.

Intel's previous research found that Loihi's search speed for feature vectors in millions of image databases is 24 times faster than CPU, and its energy consumption is 30 times lower.

-Optimization and Search

Intel has found that Loihi's efficiency in solving optimization and search problems is 1000 times higher and 100 times faster than traditional CPUs. This study can be used for real-time planning of drones and making complex navigation decisions, as well as expanding to complex data center loads to assist in tasks such as train scheduling and logistics optimization.

-Robotics technology

Researchers from Rutgers University and Delft Institute of Technology have demonstrated the application of robot navigation and micro drone control on Loihi.

The drone at Delft University of Technology uses a pulse network consisting of 35 neurons that can evolve for optical flow landing, with a frequency exceeding 250KHz.

Rutgers University found that under the same performance, the power consumption of the Loihi solution is 75 times lower than that of traditional mobile GPUs.

Loihi can also successfully learn many OpenAI Gym tasks, with accuracy comparable to that of Deep Actor Networ, and energy consumption 140 times lower than mobile GPU solutions.

Intel also demonstrated how Loihi can adaptively control a horizontal tracking drone platform, achieving a closed-loop speed of up to 20KHz and a visual processing delay of 200 microseconds, which is 1000 times higher than traditional solutions.

By the way, the name "Loihi" comes from a continuously erupting active volcano on the seabed of Hawaii, each eruption expanding the scope of the island. Intel named the neuromorphic chip after this, hoping that it can provide more powerful artificial intelligence capabilities through continuous self-learning.

Prev: Researchers evaluate the application of III-V solar cells in electric vehicles