Robust framework for intelligent neural interface systems

1 . A neural interface system, comprising:

feature extraction circuitry to identify one or more features of one or more input signals; and

neural processing circuitry, the neural processing circuitry configured to:

identify a current context of a plurality of contexts based on a first trigger event;

decode the one or more features of the one or more input signals to determine a first task of a plurality of tasks in the current context;

responsive to detecting a second trigger event, change the current context to a next context of the plurality of contexts; and

build a neural network decoder directly from raw data, wherein build the neural network decoder further comprises:

learn the one or more features of a deep neural network directly from the raw data; and

adjust the one or more features and the neural network decoder based on a predicted output during normal use, wherein adjusting the features and the neural network decoder during normal use does not require recalibration by a user.

2 . The neural interface system of claim 1 , wherein the first trigger event and the second trigger event each includes at least one of the one or more features of a second input signal, a location, a button press, a manual signal, a brainwave electrical signal, training data, and/or a training routine associated with a context.

3 . The neural interface system of claim 1 , wherein the neural processing circuitry is further configured to:

update one or more model parameters of the neural processing circuitry for any context of the plurality of contexts using new data and predicted output during usage of the current context, wherein the one or more model parameters of a neural network decoding model are updated for all previously collected contexts.

4 . The neural interface system of claim 1 , wherein the feature extraction circuitry is configured to identify the one or more features of a brain-computer interface (BCI) signal.

5 . The neural interface system of claim 1 , wherein the feature extraction circuitry is configured to identify the one or more features of an electromyography (EMG) signal.

6 . The neural interface system of claim 1 , wherein the neural processing circuitry further comprises:

a long short-term memory (LSTM) neural processing layer to identify patterns in sequences of the one or more input signals;

a convolution neural network layer in which input signals and feedback data are assigned to one or more nodes of the convolution neural network layer and selected features are assigned weighting factors; and

a fully connected layer comprising a single-layer neural network architecture.

7 . The neural interface system of claim 1 , further comprising channel mask layer circuitry to exclude selected input signals.

8 . The neural interface system of claim 1 , further comprising transformation circuitry to transform the one or more input signals from a first type of signal to a second type of signal.

9 . A non-transitory storage device that includes machine-readable instructions that, when executed by one or more processors, cause one or more processors to perform operations comprising:

identify a current context of a plurality of contexts based on a first trigger event;

decode one or more features of one or more input signals to determine a first task of a plurality of tasks in the current context;

responsive to detecting a second trigger event, change the current context to a next context of the plurality of contexts; and

build a neural network decoder directly from raw data, wherein the neural network decoder further comprises:

learn the one or more features of a deep neural network directly from the raw data; and

adjust the one or more features and the neural network decoder based on a predicted output during normal use, wherein adjusting the features and the neural network decoder during normal use does not require recalibration by a user.

10 . The non-transitory storage device of claim 9 , wherein the first trigger event and the second trigger event each includes at least one of the one or more features of a second input signal, a location, a button press, a manual signal, a brainwave electrical signal, training data, and/or a training routine associated with a context.

11 . The non-transitory storage device of claim 9 , wherein the machine-readable instructions that, when executed by the one or more processors, cause the one or more processors to perform the operations, further comprises:

update one or more model parameters of a neural network decoding model for any context of the of the plurality of contexts using new data and predicted output during usage of the current context, wherein the one or more model parameters of a neural network decoding model are updated for all previously collected contexts.

12 . The non-transitory storage device of claim 9 , wherein the machine-readable instructions that, when executed by the one or more processors, cause the one or more processors to perform the operations, comprising:

identify the one or more features of an electromyography (EMG) signal.

13 . The non-transitory storage device of claim 9 , wherein the machine-readable instructions that, when executed by the one or more processors, cause the one or more processors to perform the operations, comprising:

exclude selected input signals.

14 . A method for selecting neural network architectures, the method comprising:

determine a directed acyclic graph (DAG) representing all potential architecture choices to search;

determining a first loss function applied to a set of training data using the DAG;

determining a second loss function applied to one or more sets of validation data using the DAG;

optimizing a plurality of neural network parameters based on the first loss function applied to the set of training data;

optimizing one or more architecture graph connections based on the second loss function applied to the one or more sets of validation data; and

determining a robust architecture based on a result of optimizing the one or more architecture graph connections based on the second loss function applied to the one or more sets of validation data.

15 . The method of claim 14 , wherein optimizing the plurality of neural network parameters based on the first loss function applied to the set of training data further comprises:

descending a gradient of the plurality of neural network parameters; and

updating a weight based on the gradient for each neural network parameter of the plurality of neural network parameters.

16 . The method of claim 14 , wherein optimizing the one or more architecture graph connections based on the second loss function applied to the one or more sets of validation data further comprises:

descending a gradient of the one or more architecture graph connections; and

updating a weight based on the gradient for each of the one or more architecture graph connections.