Hi @ManChild001,
CNN conversion is described in our PCT patent application WO2020092691
https://worldwide.espacenet.com/patent/search/family/070458523/publication/WO2020092691A1?q=WO2020092691 .
The production of the Akida chip was delayed for a year to add the CNN2SNN conversion capability because most installed image detectors talk CNN (again 20th century von Neumann data structure), but processing in SNN is much mor efficient.
https://worldwide.espacenet.com/patent/search/family/074189798/publication/WO2021016544A1?q=WO2021016544
The latest published PCT patent application WO2021016544 (I've referred to this patent application so many times, I remember the number without looking it up) describes the learning feature which [relates to/differentiates Akida from?] the Masqualier article on back-propagation @Fact Finder posted above in an attempt to destroy my few remaining grey cells. (Apologies to M Poirot). [Little does he know that they have been preserved in alcohol!]
As far as I can make out (I'm guessing), the filter (a scanning block which sequentially examines a group of adjacent pixel values across the whole of the input data frame) takes some account of the adjacent pixel values of the specific pixel value being assessed to approximate the columnar structure as shown at Figure 9 (page 6) in: How BrainChip is changing the AI Industry:
https://brainchipinc.com/wp-content/uploads/2020/05/BrainChip_tech-brief_6-How-BrainChip-is-Changing-AI_v1.2.pdfwhich someone posted the other day.
Figure 9: Neurons in the brain are organized in columns that trigger each other like dominoes, resulting in sequential memory and prediction of the next sensory perception a few milliseconds before it arrives. Prediction is an important facilitator of learning in the brain that has been overlooked for more than 30 years.
...
The neocortex is organized into vertical columns called cortical columns. Cortical columns consist of dense vertical connections between neurons within a single column. Cortical columns also feature lateral neuronal connections to neighboring columns (Figure 9). [#### is this what the filter does?####]
...
When a neuron partially depolarizes a neuron in an adjacent column, it puts that neuron into a predictive state. When sensory stimuli arrive a few milliseconds later, the prediction is either confirmed or denied. If the prediction is false, connections are weakened through an inhibitory response. If on the other hand the prediction is confirmed, the connection between columns are strengthened by increasing weight values. This process, known as Spike Time Dependent Plasticity (STDP), continues to
form linked lists of cortical columns, forming sequence memory. BrainChip believes that this process exists everywhere in the cortex, including the visual cortex, the sensory-somatic cortex, and the prefrontal cortex.
...
T
he company eventually expects machine learning to evolve into a new artificial neural network architecture, a spiking cortical network, that is superior to current CNN and SNN architectures and appropriate for industrial application in visual object learning and classification, big data analysis, feature extraction and classification, industrial automation, and robotics. A spiking cortical neural network is capable of real-time learning, and will require significantly less training time than a CNN.
To me the process of depolarizing adjacent neurons is analogous to cross-talk on untwisted phone line pairs. (A varying electrical curent produces an accompanying field which can be picked up by nearby conductor.) I have a feeling that the 3*3 or 5*5 filter of WO2021016544 can provide a means of taking the "crosstalk" between neurons into account, but that is speculation.
It should be noted from the timeline in Figure 8, this paper is a few years old, starting at 2018 and there has been a bit of slippage, with the Akida 2 test chip set down for 2020. However, we know that there have been major changes to Akida 1 in the meantime.
We also "know" (or think we do) that the Akida 2 design is well nigh complete.
(20min delay)