New MRAM Memory Will Accelerate A.I. and ML
Just worth mentioning for those who wonder where breakthroughs could come from in the near future.
A company in Fremont, Ca. (The home of many chip companies and Tesla Motors) called ‘Spin Memory’ has developed a ‘Magnetoresistive Access Memory’ technology that is quite interesting.
They use the stochastic behavior of using a magnetic element when switching between 1 and 0 to in essence, reduce the footprint of a memory cell by an order of magnitude over Static RAM. If true, it also reduces the power consumption since it uses has zero power to retain its contents when turned off. The MRAM is works like a flash static ram with much higher density and much lower power consumption, making it a threat to Flash memory, processor cache and mobile device memory uses.
What about Deep Learning?
Stochastic gradient descent plays a major role in deep learning and image recognition. Spin created a binarized neural network, similar to the Xnor.ai mentioned in another article on this site. Even with the injection of errors into the Binarized Neural Network reaching 30%, they were still seeing comparable accuracies to CIFAR 10.
Here is a good reading list to come up to speed :
Great explainer article for Spin Memory
Spin Memory to Discuss the Energy Catastrophe of Artificial Intelligence Applications at EMAI Workshop 2019
This writeup should give you a better idea why Spin Memory is looking very disruptive in the near future.