Competitive landscape, page-433

Hi ui,

I have your interlocutor on ignore, and we did look at Syntiant some time ago.

At the risk of repetition, remember that Dr Frankenstein has not retired:

US2019034790A1 Systems And Methods For Partial DigitalRetraininghttps://worldwide.espacenet.com/patent/search/family/065138355/publication/US2019034790A1?q=US2019034790A1

[0007] Disclosed herein is a neuromorphic integrated circuitincluding, in some embodiments, a multi-layered analog-digital hybrid neural network. The neural network includes a number of analog layers configured to include synaptic weights between neural nodes of the neural network for decision making by the neural network. The neural network also includes at least one digital layer. The digital layer is configured for programmatically compensating for weight drifts of the synaptic weights of the neural network, thereby maintaining integrity of the decision making by the neural network.

[0056] FIG. 5 illustrates a multi-layered hybridanalog-digital neural network 500 in accordance with some embodiments. Asshown, the hybrid neural network 500 includes a number of data inputs, a numberof analog layers, a digital layer, and a number of data outputs. The numberof analog layers is disposed between the number of data inputs and the digitallayer. The digital layer is disposed between the number of analog layers and anumber of data outputs. Programmable cells including transistors (e.g., cellsincluding transistors M 1 and M 2 of FIG. 4) in the number of analog layers canbe programmed with an initial set of weights as set forth herein for one ormore classification problems, one or more regression problems, or a combinationthereof. During operation of the number of analog layers, which are disposed inan analog multiplier array (e.g., the analog multiplier array 300 ), theweights are multiplied by input currents to provide output currents that arecombined to arrive at a decision of the hybrid neural network 500 by means ofone or more of the number of data outputs. Decision making for the regressionproblems includes predicting continuous quantities corresponding to data inputinto the number of data inputs (e.g., estimating a person's age from aphotograph). Decision making for the classification problems includespredicting discrete classes corresponding to data input into the number of datainputs (e.g., classifying an image as an image of a cat or a dog).

I'm not even sure they can make a hybrid analog-digital NN on a single wafer.