Information Systems Laboritories, page-87

Hi FF,

In the text, 26 May 2021 CEO's Address, our interim CEO said:

https://hotcopper.com.au/data/announcements/ASX/2A1300009_BRN.pdf

The AKD1000 chip is in production and we expect sales to take off soon after. Initially, we expect to sell complete modules while our clients are developing their own products and boards. Those modules will contain the AKD1000 chip and the MetaTF tools. This is where we expect the first production chips to be absorbed.

Sales are expected to increase once clients have finished their development work and are integrating the AKD1000 into their products. Module sales are likely to continue into the future, followed by new products such as the AKD2000, which will be a chip with additional features to execute sequence learning networks, known in the industry as LSTM and Transformer networks.

Important parts of the AKD2000 are already working now in the lab here in Perth. The AKD3000 chip is in development and will be aimed at capsule networks that are under development at Google and the cortical neural networks of the future.

I already gave away the clue to this slide; Akida AKD1000 is not the end, but the beginning of an exciting range of advanced neural network products, each aimed at a specific market segment of this evolving technology.

I’ll illustrate the reason for these products in an example:
Today, if a plastic bag blows across the street, a car equipped with AI sees an object and hits the brakes, or worse, takes evasive action. Our future networks will be able to learn the difference between a plastic bag, blowing across the street and a rock by their behaviour. They will learn from sequences of events and from observing behaviour. A rock does not get blown in the wind. A door has a handle that makes it open. Objects have not only a shape, but also behaviour and a location in space. Our brain understands and predicts the behaviour of objects and sounds.

We aim to build that intelligence into future products so that we enable our clients not only to build intelligent products, but safe and beneficial products.

The slides show LSTM in Akida `1500 and Akida 2000.



So LSTM is shown in both Akida 1500 and Akida 2000.

I don't know if this is directly on point, but it illustrates how the asynchronous spikes are sorted into a packet correlated to the pixel "address".

We all remember our favourite Akida NPU patent application:

WO2020092691A1 AN IMPROVED SPIKING NEURAL NETWORK





This includes a packet register 104 which


Referring to Fig 6A, the 24 (arbitrary number) of asynchronous spikes arrive in a sequence determined by the order of firing of the neurons associated with specific neuron locations, ie, a spike addressed to neuron 1 arrives first, then a spike addressed to neuron 6, then a spike addressed to neuron 23, then another spike addressed to neuron 1, then a spike addressed to neuron19 arrives ... in spike buffer 601. Each spike has a destination neuron address (index number) attached and the packetized uses the address to sort the spikes into the order corresponding to the address of the corresponding neurons as shown in Fig 6B at 603.


So my thought is: can we simply duplicate the spike packet register (SPR) 603 in an additional SPR for one packet duration so the duplicate SPR can be compared with the subsequent packet to implement LSTM?

Or, if we need a longer term comparison. 2 or more additional SPRs can be added so a series of "snapshot" packets can be compared?