Method for signal representation and construction

1 . A Deep-Learning (DL) network system for representation and construction of input signals comprising:

an encoder stage configured to process input signals, said encoder stage comprising Dilated Convolutions (DC) and mapping of the input signals into a latent embedded representation, learning a set of features that encode said input signals, wherein learning feature representations maintain a sample size the same as said input signals;

an encoding capturing a latent space representation of said input signals in a learned feature space; and

a decoder stage mapping latent features back to constructed or reconstructed output signals of a same size as said input signals;

whereby said constructed signal output has a same dimensionality and representation as said input signals providing an alternative representation of said input signal.

2 . The system of claim 1 comprising:

a training procedure comprising:

training inputs;

training targets; and

at least one training loss function, whereby weights of said encoder stage and said decoder stage layer are updated.

3 . The system of claim 1 , wherein said encoder stage comprises six signal dilated causal convolutions (SDCC) layers comprising four residual connection layers and four skip connection layers.

4 . The system of claim 1 , wherein said encoder stage comprises six Signal Dilated Causal Convolutions (SDCC) layers comprising, from input to output, dilation rate, kernel size, and number of filters parameters of (1, 2, NF), (2, 4, 50), (8, 4, 50), (32, 4, 50), (128, 4, 50), and (362, 4, 50).

5 . The system of claim 1 , wherein latent feature maps are learned from an unsupervised signal construction task and then transferred into other supervised tasks as a base feature extraction of said network.

6 . The system of claim 1 , wherein said encoder is trained and then transferred to be used for other applications; and

wherein said system is trained with a loss function that produces gradients to determine how to update said network.

7 . The system of claim 1 , wherein a trained encoder is used for supervised learning tasks with the decoder being replaced by classification layers; and

wherein said system is trained with a loss function to minimize a difference between a signal produced by said decoder and a target training signal and a resulting output of a loss function is used to determine how to update said network.

8 . The system of claim 1 , wherein said encoder stage comprises Signal Dilated Causal Convolutions (SDCC) layers plus a REctified Linear activation function (ReLU) and Batch Normalization (BN); and

said decoder stage comprises a set of Dense, fully-connected, layers plus a REctified Linear activation function (ReLU) to map from a latent feature map space to an output signal.

9 . A non-transitory computer readable medium, having stored thereon, instructions that, when executed by a computing device, cause the computing device to perform a representation and reconstruction of digital input signals method comprising:

receiving said digital input signal in an encoder stage, wherein a set of Signal Dilated Causal Convolutions (SDCC) layers extract features from said digital input signal into a latent feature map space, wherein learning feature representations maintain a same sample size as the input;

receiving, in a decoder stage; output from said encoder stage; and

outputting said reconstruction of said input signal from said decoder stage;

whereby said reconstructed signal output has a same dimensionality and representation as said input signals.