US-12626090-B2 - Hierarchical neuromorphic sensor array with integrated learning for physicochemical property prediction

US12626090B2US 12626090 B2US12626090 B2US 12626090B2US-12626090-B2

Abstract

A modular artificial neural sensing system includes a hierarchical network of neural sensing units including a neuromimetic sensor array of artificial sensory synapses and sensory neurons for receiving physicochemical sensed signals and for outputting sensor output signals. An artificial neural network processor is adapted for processing the sensor output signals and includes processor neurons interconnected by processor synapses forming first connections and second connections. The processor outputs processor output signals. A first sensor interface feeds processed or unprocessed sensed signals into the processor. A second sensor interface receives output predicted signals from other neural sensing units and feeds processed or unprocessed output predicted signals into the processor. A signal decoder decodes the processor output signals and outputs decoder output signals. An error feedback module receives the decoder output signals and teaching signals for generating error signals depending on a difference between teaching signals and decoder output signals.

Inventors

Josep Maria MARGARIT TAULÉ
Shih-Chii Liu
Cecilia JIMÉNEZ JORQUERA

Assignees

Universität Zürich
CONSEJO SUPERIOR DE INVESTIGACIONES CIENTÍFICAS

Dates

Publication Date: 20260512
Application Date: 20210121
Priority Date: 20200206

Claims (15)

1 . A modular artificial neural sensing system comprising a hierarchical network of neural sensing units, each neural sensing unit comprising: a neuromimetic sensor array of artificial sensory synapses and sensory neurons, for receiving physicochemical sensed signals and for outputting a set of sensor output signals; an artificial neural network processor for processing the sensor output signals, the artificial neural network processor comprising artificial neural network processor neurons interconnected by artificial processor synapses forming first connections and different, second connections, the artificial neural network processor being configured to output a set of artificial neural network processor output signals; a first sensor interface for feeding processed or unprocessed physicochemical sensed signals into the artificial neural network processor of the respective neural sensing unit; a second sensor interface for receiving output predicted signals from other neural sensing units, and for feeding processed or unprocessed output predicted signals into the artificial neural network processor of the respective neural sensing unit; a signal decoder for decoding the artificial neural network processor output signals, and for outputting a set of decoder output signals, which are output signals to predict; and an error feedback module configured to receive the decoder output signals and teaching signals for generating a set of error signals configured to be fed back through a first error feedback module interface to the artificial neural network processor and the signal decoder, a respective error signal depending on a difference between a respective teaching signal and a respective decoder output signal, wherein: the first connections of the respective neural sensing unit are configured to be locally trained by at least any of the processed or unprocessed physicochemical sensed signals received from the first sensor interface and/or any of the processed or unprocessed output predicted signals received from the second sensor interface, the second connections of the respective neural sensing unit are configured to be locally trained by at least any of the error signals, and the neural sensing units are interconnected so that any of the decoder output signals from a given layer are configured to be fed at least into a second sensor interface of another neural sensing unit of a subsequent, lower layer.
2 . The neural sensing system according to claim 1 , wherein: the synapses of the neural sensing units are a first-type dynamical part, the neurons of the neural sensing units are a different, second-type dynamical part, instances of the first-type dynamical part being configured as single-input, single-output integrator synapses, and instances of the second-type dynamical part being configured as multiple-input, single-output integrator neurons, and the neurons and synapses of the neural sensing units form a directed neural network exhibiting continuous-time dynamical behaviour.
3 . The neural sensing system according to claim 1 , wherein: the signal decoder comprises a set of artificial decoder neurons forming a set of signal read-out units, and a set of artificial decoder synapses forming third connections, and the artificial decoder synapses connect the artificial neural network processor to the artificial decoder neurons.
4 . The neural sensing system according to claim 3 , wherein the artificial decoder neurons and/or the third connections are configured to be trained by at least the error signals.
5 . The neural sensing system according to claim 1 , wherein: the first connections are configured to be trained previously or simultaneously with the second connections, and at a separate time scale from the one at which learning of the second connections evolve, learning in the first connections is driven by the processed or unprocessed physicochemical sensed signals and the processed or unprocessed output predicted signals, and the physicochemical sensed signals and the output predicted signals have higher average frequencies than average frequencies of the teaching signals used to generate the error signals configured to modulate learning in the second connections.
6 . The neural sensing system according to claim 1 , wherein: the artificial neural network processor neurons are configured to be trained by at least any of the signals employed for training the first connections, or any of the error signals, and wherein the error signals are generated using the teaching signals of lower average frequency than the average frequencies of the physicochemical sensed signals and/or the output predicted signals.
7 . The neural sensing system according to claim 1 , wherein the artificial neural network processor is a subcircuit of a recurrent neural network.
8 . The neural sensing system according to claim 1 , wherein: the error feedback module is connected to the artificial neural network processor by a second error feedback module interface, and the first and second error feedback module interfaces comprise artificial error feedback synapses.
9 . The neural sensing system according to claim 1 , wherein: the first sensor interface comprises first artificial interface synapses, and the second sensor interface comprises second artificial interface synapses.
10 . The neural sensing system according to claim 1 , wherein the neural sensing units are configured to implement sparse coding of the physicochemical sensed signals and the output predicted signals by lateral inhibition between the artificial neural network processor neurons.
11 . The neural sensing system according to claim 1 , wherein the error feedback module comprises a subtractor unit configured to subtract a respective decoder output signal from a respective training signal, or vice versa.
12 . The neural sensing system according to claim 1 , wherein the error feedback module comprises an artificial neural network configured to generate the set of error signals.
13 . The neural sensing system according to claim 1 , wherein at least some of the artificial neural network processor neurons comprise the set of sensory neurons.
14 . A method of operating the neural sensing system according to claim 1 , wherein the method comprises: selecting the teaching signals; feeding the teaching signals into the error feedback module; carrying out error-unmodulated training of at least the first connections by using at least any of the signals coming from the first sensor interface, and any of the signals coming from the second sensor interface; and carrying out error-modulated training of at least the second connections by using at least any of the error signals.
15 . The method according to claim 14 , wherein the method further comprises feeding the decoder output signals into one or more other neural sensing units once the respective neural sensing unit has been trained.

Description

TECHNICAL FIELD The present invention relates to an artificial neural sensing unit for predicting dynamic physicochemical properties (i.e., dynamic properties which are physical and/or chemical). The invention equally relates to a modular and hierarchical artificial neural sensing system comprising a plurality of artificial neural sensing units. The invention also relates to a method of operating the artificial neural sensing unit. BACKGROUND OF THE INVENTION Virtual sensing uses the information available online from other sensor measurements to estimate a property of interest. This approach represents an attractive solution for providing data otherwise unfeasible to obtain due to either the inexistence of specific sensors or the high costs associated to manufacturing such technologies. Many existing virtual sensing solutions are based on software implementations of machine learning models, trained with big past empirical data collected from sensors, and running aside on high-speed clocked, synchronous computing devices. Most solutions decouple today multi-sensor measurement, artificial intelligence (AI) computation, and system memory, sequentially processing data and moving data at limited throughputs between the three blocks. This makes it difficult for such solutions to adapt to new operating conditions in real time. In addition, this limits both miniaturisation and energy autonomy to operate locally in remote and narrow sensing spots. The above challenges in mind, the engineering field of adaptive control theory provides general principles for estimating complex dynamics online from sensor observations and continuous output feedback. Resembling the parallel processing architecture of biological brains, these principles can be applied not only to generate, but also to learn the dynamics with local plasticity rules on predictive networks of neurons and synapses. Embodiments of both parts can operate dynamically in continuous time (i.e., varying at any time instant) and can have adaptive memory. Significant theoretical, computational, and experimental studies in neuroscience endorse such models of multisensory cortical processing. By using more artificial neurons than inputs to the neural network, and by feeding back predictions generated from neuronal activity (i.e., instantaneous spike firing rate), the resulting recurrent networks are able to efficiently encode task dynamics as sparse representations in space and time. This strategy is deemed advantageous for saving energy, expanding storage capacity in associative memories, and representing signals explicitly. Sparse sensor coding typically makes neurons more selective to specific patterns of input in a way that facilitates hierarchical feature extraction of multi-sensor stimulus. Because sparsely coded signals are overrepresented in the number of neural units, they are also robust to environmental constraints, such as noise and reduced connectivity between neurons. Additionally, subtracting neural network model predictions allows curtailing power consumption by encoding only the unexpected component(s) of sensor information. In the following description, sparsely coded signals are understood to represent information by using a small number of strongly active neurons out of a large population at any time instant. In non-sparse coding, feature representation is distributed simultaneously over the neurons that constitute the network. A publication entitled “Online reservoir adaptation by intrinsic plasticity for backpropagation—decorrelation and echo state learning”, Jochen J. Steil, Neural Networks, Volume 20, Issue 3, April 2007, Pages 353-364, discloses a solution that uses a biologically motivated learning rule based on neural intrinsic plasticity to optimise reservoirs of analogue neurons. However, the existing solutions still have many limitations, in particular concerning miniaturisation, energy consumption, calibration and scalability. SUMMARY OF THE INVENTION The objective of the present invention is thus to overcome at least some of the above limitations relating to sensing circuits. According to a first aspect of the invention, there is provided a modular artificial neural sensing system as recited in claim 1. The respective neural sensing unit of the system integrates a physicochemical sensing (sub)network, i.e., a set of sensors, in an encoding-decoding network, i.e., coders, together with an error feedback module. The resulting neural sensing unit is an artificial, recurrent neural (super)network operating as a dynamical model for predicting new physicochemical signals of interest, in continuous time and in a common system carrier. There is thus provided a modular, artificial neural sensing system comprising a hierarchical structure of the artificial neural sensing units. The proposed artificial neural sensing unit (NSU) or the artificial neural sensing system provides at least some of the following advantages: Miniaturisation: The sensors,