Thanks Ninja.I've been thinking for some time that neural...

Thanks Ninja.

I've been thinking for some time that neural networks could be the ultimate 'killer application' for WBT's technology.

I remember seeing a test circuit board for BRN's Akida chip that has a NOR flash chip right next to their AKD1000 chip. The function of this appears to be to store program weightings (i.e. the learning). On power up the Akida chip boots up from this flash memory.

In-memory neural processing in a ReRAM chip does away with this layer of complexity, since ReRAM is non-volatile and the learning state is simply retained directly in the cell states.

But looking ahead long-term I see a whole other angle here:

WBT have pointed out in their presentations that a WBT ReRAM cell can have a continuous range of resistance. It is possible to differentiate more than 2 resistance states, opening up the possibility of more than one bit per cell. So 4 states would give you 2 bits, 8 states 3 bits, 16 states 4 bits and so on.

So we are not necessarily restricted to binary (i.e. 2 state) cells. We can have octal (8 state) cells, perhaps even hexadecimal (16 state) cells (not sure what you would call a 4 state cell - a quadral cell?).

In the abstract (in your link) they mention that MB scale neural networks yield Tera FLOP level computation complexity. But they are talking only about binary neural networks.

Multi-state cells could vastly increase the computational complexity over binary cells. Intuition tells me that this would not be a simple linear relationship, with octal cells for example yielding an 8 x increase. I'm no mathematician, but I feel it would depend also on the level of inter-connectedness of the network such that the multiplier factor would be more like 8 raised to a power (e.g. 8^4) .

But this may go even further, and at the risk of coming across as a total kook I'll explain:

WBT have also pointed out in their presentations that the continuous (i.e. fundamentally non-numerical) nature of the resistance of the WBT ReRAM cell is functionally analogous to a neural synapse. Allowing the cells to be continuous, rather than differentiating the resistance level into discrete states, actually sounds much easier to implement at the hardware level. What it it would do for the computational complexity just boggles the mind.

I also foresee multi-layer neural networks where the layered structure is created not just to increase packing density, but to enable vertical interconnection, transforming a planer neural architecture into a truly 3D neural network.

We are now way beyond self-driving vehicles here. A tera scale, non-numerical, 3D neural network could be the brain of the Cyberdyne Systems T-800.

He-he. Killer app - get it?