2020 BRN Discussion, page-29258

Couldn't find any published patent applications for the authors with IMEC. [18 month lacuna applies].

I see they use integrate-and-fire neurons, as does Akida, while the accepted orthoxoxy uses leaky integrate-and-fire neurons:

Another aspect that, at first sight, might appear as a limitation of µBrain is the use of Integrate-and-Fire (IF) neurons. However,there is recurrent synaptic connectivity among neurons the absence of leakage in the neurons may see as unnecessarily restrictive to the effectiveness of recurrent network architectures.In practice, however, quite the opposite holds. It is easy to introduce leakage at a fine-grained neuron level (different leak functions and with varying parameters per neuron); by sacrificing for this purpose, one neuron’s inputs. This choice has been motivated by the intended use of µBrain primarily for experimental purposes.

It is not very flexible in use.

p13: In µBrain, our approach is to sacrify runtime flexibility for efficiency. Therefore, we decided to perform mapping-synthesis co-optimization. After synthesis and fabrication of the chip, in µBrain, it is only possible to modify the synaptic weights of the SNN but not the main configuration(synaptic connectivity). This saves substantial area and allows for highly efficient implementation of the processing unit for a targetapplication (for example, when integrating with a radar sensor).

...

An interesting point is that most of IMEC's earlier NN patents use analog, whereas this document points out weaknesses of analog while paying lip service to the party line.

p13: Moradi presented an analog neuromorphic processor. Even though the analog design has clear advantages over the digital one, it is not easily integratable and synthesizable with other digital units (e.g., sensors) and therefore different from our proposed solution. As we discussed before, analog design is also vulnerable to manufacturing variations, making its simulation and training in software difficult. It is challenging to use for critical applications like healthcare. Nevertheless, µBrain gets as close as possible to an analog design by featuring a clock-less architecture (truly event-driven) and co-localizing computation and memory in the same die.

In Table 2, Davis et al is, of course, our friends from Intel.