Methods for compressing a neural network

1 . A method for compressing a neural network, comprising:

executing the neural network by a plurality of member vehicles of a vehicle fleet, wherein each member vehicle comprises a local memory and the vehicle processor, wherein the local memory stores the neural network and the vehicle processor executes the neural network locally;

determining, during at least one inference phase by the member vehicles, a selection of elements of the neural network that should be pruned, wherein determining the selection of elements comprises:

feeding the neural network temporally adjacent input data;

calculating temporal activation differences for elements of the neural network, wherein a temporal activation difference for an element represents a change in output activation of the element between the temporally adjacent input data;

ranking the elements according to their respective temporal differences; and

selecting the elements having temporal activation differences exceeding a predefined threshold value;

transmitting, by the member vehicles of the fleet, the respectively determined selection to a central server;

transmitting, by the member vehicles of the fleet, rankings of the selection of elements;

providing, by the central server, a merged ranking on the basis of the transmitted rankings;

merging, by the central server, the respectively transmitted selections and generating a merged selection on the basis of the merged ranking;

pruning, by the central server, the neural network on the basis of the merged selection, wherein elements most frequently selected by the member vehicles of the vehicle fleet are pruned;

transmitting, by the central server, the pruned neural network to at least one member vehicle of the vehicle fleet;

storing, by the at least one member vehicle of the vehicle fleet, the pruned neural network in the local memory of the member vehicle, which pruned neural network replaces the neural network in the local memory; and

using the pruned neural network in one or more of a driver assistance system and a driving function for automated driving of the respective at least one member vehicle.

2 . The method of claim 1 , wherein the selection is transmitted to the central server if at least one transmission criterion is met.

3 . The method of claim 2 , comprising retraining the neural network after the pruning.

4 . The method of claim 2 , comprising deactivating the elements for pruning at least temporarily.

5 . The method of claim 1 , comprising retraining the neural network after the pruning.

6 . The method of claim 1 , comprising deactivating the elements for pruning at least temporarily.

7 . The method of claim 1 , wherein the pruning only takes place if at least one trigger criterion is met.

8 . A system for compressing neural network, comprising:

a plurality of member vehicles of a vehicle fleet and a central server;

wherein each member vehicle comprises:

a local memory; and

a vehicle processor;

wherein the vehicle processor is configured to:

execute the neural network stored in the local memory;

determine, during at least one inference phase, a selection of elements of the neural network that should be pruned, wherein determining the selection of elements comprises:

feeding the neural network temporally adjacent input data;

calculating temporal activation differences for elements of the neural network, wherein a temporal activation difference for an element represents a change in output activation of the element between the temporally adjacent input data;

ranking the elements according to their respective temporal differences; and

selecting the elements having temporal activation differences exceeding a predefined threshold value;

transmit the determined selection to the central server;

transmit rankings associated with the selected elements to the central server;

receive, from the central server, a pruned neural network;

store the pruned neural network in the local memory of the member vehicle, wherein pruned neural network replaces the neural network in the local memory; and

provide the pruned neural network to one or more of a driver assistance system and a driving function for automated driving of the member vehicle;

wherein the central server comprises:

a processor and a memory;

wherein the memory comprises computer instructions that causes the processor to:

receive selections of elements of the neural network from the plurality of member vehicles of a vehicle fleet;

receive rankings of the selection of elements from the plurality of member vehicles of the vehicle fleet;

provide a merged ranking on the basis of the received rankings;

merge selections of elements of the-a neural network transmitted by the member vehicles of the vehicle fleet on the basis of the merged ranking;

generate a merged selection;

prune the neural network on the basis of the merged selection, wherein the elements that are most frequently selected by the member vehicles of the vehicle fleet are pruned; and

transmit the pruned neural network to at least one member vehicle of the vehicle fleet.