I know is an older article but edge impuls posted it today
Edge Impulse @EdgeImpulse
45min
.
@CEA_Leti researchers are coupling innovative sensors with RRAM–based neuromorphic computation to build ultra-low-power systems for edge AI applications.
I see this as the K-Mart, or cheaper, version of achieving some of the benefits of neuromorphic computing. The cheapness is in the cost of RRAM and power savings are provided by the low power consumption of RRAM. I doubt it has one-shot learning nor uses spikes or sparcity. It could be used to implement LTSM however, with a lot of effort by each individual developer.
RRAM-based neuromorphic computation will have its uses and will be useful in places where Akida offers too much functionality. Mainly purpose-built smarts for a specific task that is trained once and doesn’t change. Hence not a competitor in my mind. And who knows, developers may play with this and consider Akida when they soon reach the limitations of RRAM based neurons.
But as with anything that is cheaper, the savings may be artificial, as more $ will need to be spent developing solutions as much of the work will need to be coded.
Maybe a RRAM based neuron, or even neural network, may work to pre-process inputs into an Akida device. Maybe even help with LSTM by persistently remembering states and weights for a few iterations (and potentially forever). This, along with persistent data storage, is one of the reasons I hope for cooperation between BrainChip and Weebit. A ball I hopefully have already started rolling.
Anything advancing the cause of neuromorphic computing is positive in my mind. Believe it or not, there still are A vast number of WANCAs out there Who need the info spoon fed to them.