US 12,229,650 B2
System and method for compact and efficient sparse neural networks
Eli David, Tel Aviv (IL); and Eri Rubin, Ma'ale Ha'hamisha (IL)
Assigned to NANO DIMENSION TECHNOLOGIES, LTD., Ness Ziona (IL)
Filed by Nano Dimension Technologies, Ltd., Ness Ziona (IL)
Filed on Feb. 13, 2023, as Appl. No. 18/168,016.
Application 18/168,016 is a continuation of application No. 16/524,571, filed on Jul. 29, 2019, granted, now 11,580,352.
Application 16/524,571 is a continuation of application No. 16/041,497, filed on Jul. 20, 2018, granted, now 10,366,322, issued on Jul. 30, 2019.
Application 16/524,571 is a continuation of application No. PCT/IL2018/051062, filed on Sep. 20, 2018.
Application PCT/IL2018/051062 is a continuation of application No. 16/041,497, filed on Jul. 20, 2018, granted, now 10,366,322, issued on Jul. 30, 2019.
Claims priority of provisional application 62/569,033, filed on Oct. 6, 2017.
Prior Publication US 2023/0196061 A1, Jun. 22, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 3/02 (2006.01); G06N 3/04 (2023.01)
CPC G06N 3/02 (2013.01) [G06N 3/04 (2013.01)] 24 Claims
OG exemplary drawing
 
1. A method for efficiently storing a neural network, the method comprising:
receiving a dense neural network comprising a plurality of weights, each weight representing a unique connection between a pair of a plurality of artificial neurons in different layers of a plurality of neuron layers, wherein the dense neural network comprises at least one hidden layer, wherein a majority of pairs of neurons in different neuron layers are connected by weights in the dense neural network;
transforming the dense neural network into the sparse neural network during a training phase so that a minority of pairs of neurons in different neuron layers are connected by weights in the sparse neural network;
storing only non-zero weights of the sparse neural network that represent unique connections between the pairs of artificial neurons and not storing zero weights of the sparse neural network that represent no connections between the pairs of artificial neurons.