US 12,468,933 B2
Efficient simultaneous inference computation for multiple neural networks
Dirk Staneker, Tübingen (DE); Hans-Georg Horst, Leonberg, GA (US); Nicolai Wacker, Flein (DE); Thomas Matschke, Bietigheim-Bissingen (DE); and Wolfgang Dressler, Vaihingen/Enz (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Mar. 4, 2021, as Appl. No. 17/192,250.
Claims priority of application No. 102020203047.2 (DE), filed on Mar. 10, 2020.
Prior Publication US 2021/0287087 A1, Sep. 16, 2021
Int. Cl. G06N 3/08 (2023.01); G06N 3/045 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/045 (2023.01)] 10 Claims
OG exemplary drawing
 
1. A method for an inference computation of a plurality of neural networks on a hardware platform, each of the neural networks including a respective plurality of neurons, each of the neurons respectively aggregates inputs into a respective neural network input using a respective transfer function characterized by weights and processes the network input into an activation using an activation function, the method comprising the following steps:
identifying a respective neuron in each respective neural network of at least two different neural networks of the networks which exists: (i) in each of the at least two different neural networks in a same form or (ii) in a form that is similar in each of the at least two different neural networks according to a predefined criterion;
combining the identified respective neurons together to form a respective unit;
simultaneously performing inference computations of the at least two different neural networks, including performing a single inference computation for the respective unit for all of the at least two different neural networks on the hardware platform instead of performing a separate inference computation for each of the identified respective neurons of the at least two different neural networks, the respective unit providing a result of the single interference computation to all of the at least two different neural networks as a set of outputs, and each respective neural network of the at least two different neural networks processing the set of outputs as an output of the identified respective neuron of the respective neural network, wherein the performed inference computations of each of the at least two different neural networks provides a respective inference result;
wherein a respective input of each the at least two different neural networks originates from one or more sensors of a vehicle, the inputs include: image data and/or video data and/or radar data and/or ultrasonic data and/or LIDAR data, wherein the at least two different neural networks are respectively trained for classifying different objects, and wherein the respective inference result of each respective neural network of the at least two different neural networks includes a classification of an object detected in the respective input of the respective neural network;
deriving an action for a driving assistance system of the vehicle or of a system for an at least partially automated control of the vehicle, based on the respective inferences results;
forming a control signal based on the derived action; and
controlling, using the control signal, the vehicle in an at least partially automated manner.