The problem is reading this patent specification is like walking over the Klondike with a metal detector - there's gold whever way you turn.
It is essential to have the drawings
https://worldwide.espacenet.com/patent/drawing?channel=espacenet_channel-459fa4b2-cfb2-4180-98b3-89f733a5f0e0and the description
https://worldwide.espacenet.com/patent/search/family/074189798/publication/WO2021016544A1?q=brainchipto hand to see what is going on.
There are major changes to the circuit layout.
Some of the changes which stood out for me:
There are "ping-pong buffers" and parallel neural processing engines in Fig 5, the scratchpad memories used in the NPUs to store neuron potentials (see the shaded squares at 760 in Figure 7.
There are 3 parallel filters (1 per input channel) in Fig 13.
[0081]
Each neuron has a particular portion of an input for which it has an interest, and therefore it activates only when a spike is within its zone of interest.
...
[0108] A filter is a 3D collection of weights (shown in blue) while a neuron is the specific application of a single filter to a specific spatial location of the inputs. A single filter will be applied to every spatial location of the inputs and generate an entire plane of neuron potential entries.
[0109] ... Neurons can be defined as filters applied to a specific x,y location of the input as shown at the bottom half. A single neuron has a single entry in the potential array shown in light orange. Each plane of potentials array is generated when a single filter is applied (centered) at every spatial location of the inputs. The single neuron entry in gray in the potential array was calculated from filter 1’s weights applied to the location of the input denoted in the lower left display of the input. The dotted red line denotes the receptive field of the neuron centered at the event. (see Fig 13)
My interpretation of this is that, in the 3*3 filters of Fig 13, the central neuron is associates with the adjacent 8 neurons in each filter.
The filter scans across each input value, making, I suppose, appropriate adjustment for edge input cells.
There are also important changes to the learning rules, wherein
[0110] ... The novel unsupervised learning rule only allows one neuron per filter to learn. An additional modification restricts the number of neurons in a layer that can learn to just one [
because all neurons for a single filter share a single set of filter weights [0111]]
; (2) during the weight-swapping, the input events are filtered to include only those events relevant to the specific neuron undergoing learning; and (3) the algorithm now supports pseudo-supervised learning to allow neurons to learn different patterns[049] ... event-based convolution is implemented in a Spiking Neural Network (SNN) using event-based rank-coding rather than a perceptron, which has advantages in speed and considerably lower power consumption. Rank coding differs from rate-coding of spike events in that values are encoded in the order of spikes transmitted. In rate coding the repetition rate of spikes transmitted expresses a real number
A technique to prevent redundant learning by different groups of neuron filters learning the same pattern by the use of input event packet labels is also provided [0115].
(20min delay)