Neuromorphic algorithm for rapid online learning and signal restoration
A computer-implemented method of training a neural network to recognize sensory patterns includes obtaining input data, preprocessing the input data in one or more preprocessors of the neural network, and applying the preprocessed input data to a core portion of the neural network. The core portion of the neural network includes a plurality of principal neurons and a plurality of interneurons, and is configured to implement a feedback loop from the interneurons to the principal neurons that supports persistent unsupervised differentiation of multiple learned sensory patterns over time. The method further includes obtaining an output from the core portion, and performing at least one automated action based at least in part on the output obtained from the core portion. The neural network may be adaptively expanded to facilitate the persistent unsupervised differentiation of multiple learned sensory patterns over time by incorporating additional interneurons into at least the core portion.
1 . A computer-implemented method of training a neural network to recognize sensory patterns, the method comprising:
obtaining input data utilizing a processor-based data sampling loop;
preprocessing the input data in one or more preprocessors of the neural network;
applying the preprocessed input data to a core portion of the neural network, the core portion of the neural network comprising a plurality of principal neurons and a plurality of interneurons, the core portion of the neural network implementing a processor-based feedback loop between the interneurons and the principal neurons that supports persistent unsupervised differentiation of multiple learned sensory patterns over time;
storing in the core portion of the neural network synaptic weights for interconnections between at least a subset of the principal neurons and at least a subset of the interneurons;
executing multiple cycles of the processor-based feedback loop of the core portion of the neural network for the applied preprocessed input data within each cycle of the processor-based data sampling loop utilized in obtaining the input data;
adjusting at least a portion of the stored synaptic weights in each of the multiple cycles of the processor-based feedback loop as part of a learning process of the core portion of the neural network;
obtaining an output from the core portion of the neural network; and
performing at least one automated action based at least in part on the output obtained from the core portion of the neural network;
wherein the neural network comprises a spiking neural network (SNN) configured to provide spike timing for the processor-based feedback loop with inhibition delaying the spike timing and excitation advancing the spike timing, and further wherein the processor-based feedback loop is configured to adapt one or more of the synaptic weights of the core portion of the neural network, and operation of the processor-based feedback loop in adapting one or more of the synaptic weights is controlled based at least in part on spike timing information determined at least in part from relative timing of spikes for at least a portion of the principal neurons and the interneurons; and
wherein the method is performed by at least one processing device comprising a processor coupled to a memory.
2 . The method of claim 1 wherein at least a subset of the principal neurons of the core portion of the neural network are configured to represent respective mitral cells of an olfactory learning system and at least a subset of the interneurons of the core portion of the neural network are configured to represent respective granule cells of the olfactory learning system.
3 . The method of claim 1 further comprising adaptively expanding the neural network by incorporating additional interneurons into to one or more of:
(i) at least one of the one or more preprocessors of the neural network; and
(ii) the core portion of the neural network.
4 . The method of claim 3 wherein the additional interneurons are incorporated into the core portion of the neural network in a manner that does not disrupt existing learned sensory patterns of the core portion of the neural network.
5 . The method of claim 1 wherein obtaining input data comprises obtaining the input data from one or more sensors.
6 . The method of claim 1 wherein a given one of the one or more preprocessors of the neural network comprises a plurality of input nodes each adapted to receive input data associated with a different data source.
7 . The method of claim 6 wherein the input nodes are adapted to receive input data from respective different sensors.
8 . The method of claim 6 wherein the given preprocessor of the neural network comprises a heterogeneous duplication preprocessor configured to statistically regularize diverse sensory inputs of the obtained input data.
9 . The method of claim 6 wherein the given preprocessor of the neural network further comprises for a particular one of the input nodes:
a plurality of excitatory feed-forward interneurons each coupled to the particular input node; and
a plurality of principal neurons each coupled to one or more of the excitatory feed-forward interneurons.
10 . The method of claim 1 wherein the core portion of the neural network comprises a synaptic interaction matrix of the principal neurons and the interneurons in which an n-dimensional representation in the principal neurons is mapped to an m-dimensional representation in the interneurons, where m>>n.
11 . The method of claim 1 wherein the neural network further comprises a neuromodulatory dynamic state trajectory configured to adjust neuronal properties systematically and select a particular outcome.
12 . The method of claim 1 wherein the neural network further comprises an inference network arranged between the principal neurons and the interneurons and configured to deliver input to the interneurons that influences how the interneurons affect the principal neurons such that the principal neurons thereby exert different effects on the interneurons and the inference network.
13 . The method of claim 1 wherein the processor-based feedback loop is configured to control delivery of synaptic inhibition information from the interneurons of the core portion back to the principal neurons of the core portion based at least in part on synaptic excitatory information delivered from the principal neurons to the interneurons.
14 . The method of claim 1 wherein the one or more preprocessors of the neural network comprise at least one layer that includes a plurality of neurons of a first type and a plurality of neurons of a second type different than the first type, and further wherein at least a subset of the neurons of the second type are configured to inhibit at least a subset of the neurons of the first type.
15 . The method of claim 14 wherein said at least one layer is configured to represent at least one glomerular layer of an olfactory learning system, the neurons of the first type are configured to represent respective external tufted (ET) cells of the olfactory learning system, and the neurons of the second type are configured to represent respective periglomerular (PG) cells of the olfactory learning system.
16 . The method of claim 14 wherein the inhibition of at least a subset of the neurons of the first type by at least a subset of the neurons of the second type comprises a graded lateral inhibition.
17 . The method of claim 1 wherein the one or more preprocessors of the neural network comprise at least one concentration tolerance preprocessor configured to limit concentration-specific variance in outputs generated in response to respective different instances of input data.
18 . The method of claim 1 wherein the one or more preprocessors of the neural network comprise at least one sensor scaling preprocessor configured to rescale outputs of multiple heterogeneous sensors such that corresponding inputs to the core portion of the neural network are statistically similarly scaled.
19 . The method of claim 1 further comprising adding one or more neurons to the neural network without disrupting an existing learned pattern obtained through previous training of the neural network.
20 . The method of claim 1 wherein the neural network is configured to implement online learning in which one or more new patterns are learned with a size of the neural network dynamically expanded, relative to a previous size of the neural network, without impairing one or more previous learned patterns of the neural network.
21 . A system comprising:
at least one processing device comprising a processor coupled to a memory;
the processing device being configured:
to obtain input data utilizing a processor-based data sampling loop;
to preprocess the input data in one or more preprocessors of a neural network;
to apply the preprocessed input data to a core portion of the neural network, the core portion of the neural network comprising a plurality of principal neurons and a plurality of interneurons, the core portion of the neural network implementing a processor-based feedback loop between the interneurons and the principal neurons that supports persistent unsupervised differentiation of multiple learned sensory patterns over time;
to store in the core portion of the neural network synaptic weights for interconnections between at least a subset of the principal neurons and at least a subset of the interneurons;
to execute multiple cycles of the processor-based feedback loop of the core portion of the neural network for the applied preprocessed input data within each cycle of the processor-based data sampling loop utilized in obtaining the input data;
to adjust at least a portion of the stored synaptic weights in each of the multiple cycles of the processor-based feedback loop as part of a learning process of the core portion of the neural network;
to obtain an output from the core portion of the neural network; and
to perform at least one automated action based at least in part on the output obtained from the core portion of the neural network;
wherein the neural network comprises a spiking neural network (SNN) configured to provide spike timing for the processor-based feedback loop with inhibition delaying the spike timing and excitation advancing the spike timing, and further wherein the processor-based feedback loop is configured to adapt one or more of the synaptic weights of the core portion of the neural network, and operation of the processor-based feedback loop in adapting one or more of the synaptic weights is controlled based at least in part on spike timing information determined at least in part from relative timing of spikes for at least a portion of the principal neurons and the interneurons.
22 . The system of claim 21 wherein the processing device is further configured to adaptively expand the neural network by incorporating additional interneurons into one or more of:
(i) at least one of the one or more preprocessors of the neural network; and
(ii) the core portion of the neural network.
23 . The system of claim 21 wherein the core portion of the neural network comprises a synaptic interaction matrix of the principal neurons and the interneurons in which an n-dimensional representation in the principal neurons is mapped to an m-dimensional representation in the interneurons, where m>>n.
24 . The system of claim 21 wherein at least a subset of the principal neurons of the core portion of the neural network are configured to represent respective mitral cells of an olfactory learning system and at least a subset of the interneurons of the core portion of the neural network are configured to represent respective granule cells of the olfactory learning system.
25 . The system of claim 21 wherein the neural network further comprises a neuromodulatory dynamic state trajectory configured to adjust neuronal properties systematically and select a particular outcome.
26 . The system of claim 21 wherein the neural network further comprises an inference network arranged between the principal neurons and the interneurons and configured to deliver input to the interneurons that influences how the interneurons affect the principal neurons such that the principal neurons thereby exert different effects on the interneurons and the inference network.
27 . The system of claim 21 wherein the processor-based feedback loop is configured to control delivery of synaptic inhibition information from the interneurons of the core portion back to the principal neurons of the core portion based at least in part on synaptic excitatory information delivered from the principal neurons to the interneurons.
28 . A computer program product comprising a non-transitory processor-readable storage medium having stored therein program code of one or more software programs, wherein the program code, when executed by at least one processing device comprising a processor coupled to a memory, causes the processing device:
to obtain input data utilizing a processor-based data sampling loop;
to preprocess the input data in one or more preprocessors of a neural network;
to apply the preprocessed input data to a core portion of the neural network, the core portion of the neural network comprising a plurality of principal neurons and a plurality of interneurons, the core portion of the neural network implementing a processor-based feedback loop between the interneurons and the principal neurons that supports persistent unsupervised differentiation of multiple learned sensory patterns over time;
to store in the core portion of the neural network synaptic weights for interconnections between at least a subset of the principal neurons and at least a subset of the interneurons;
to execute multiple cycles of the processor-based feedback loop of the core portion of the neural network for the applied preprocessed input data within each cycle of the processor-based data sampling loop utilized in obtaining the input data;
to adjust at least a portion of the stored synaptic weights in each of the multiple cycles of the processor-based feedback loop as part of a learning process of the core portion of the neural network;
to obtain an output from the core portion of the neural network; and
to perform at least one automated action based at least in part on the output obtained from the core portion of the neural network;
wherein the neural network comprises a spiking neural network (SNN) configured to provide spike timing for the processor-based feedback loop with inhibition delaying the spike timing and excitation advancing the spike timing, and further wherein the processor-based feedback loop is configured to adapt one or more of the synaptic weights of the core portion of the neural network, and operation of the processor-based feedback loop in adapting one or more of the synaptic weights is controlled based at least in part on spike timing information determined at least in part from relative timing of spikes for at least a portion of the principal neurons and the interneurons.
29 . The computer program product of claim 28 wherein the program code when executed further causes the processing device to adaptively expand the neural network by incorporating additional interneurons into one or more of:
(i) at least one of the one or more preprocessors of the neural network; and
(ii) the core portion of the neural network.
30 . The computer program product of claim 28 wherein the core portion of the neural network comprises a synaptic interaction matrix of the principal neurons and the interneurons in which an n-dimensional representation in the principal neurons is mapped to an m-dimensional representation in the interneurons, where m>>n.