Autoencoder Linearization

The autoencoder is first trained to reconstruct its own input—the PA's output signal—without requiring a desired linear reference. This self-supervised reconstruction forces the bottleneck layer to learn a compressed latent representation that captures the essential nonlinear and memory characteristics of the amplifier. Only after this phase is the decoder replaced or fine-tuned for the predistortion mapping.

The constricted middle layer of the autoencoder acts as an information bottleneck, forcing the network to discard noise and redundant signal components while preserving the manifold of PA distortion. This learned latent space provides a robust, lower-dimensional feature set that serves as an effective initialization for the predistorter, reducing sensitivity to measurement noise in the training data.

By training the autoencoder to reconstruct clean signals from intentionally corrupted inputs—a technique known as a Denoising Autoencoder (DAE)—the model learns to separate the fundamental PA nonlinearity from stochastic thermal noise and measurement artifacts. This produces a predistorter that is inherently more robust to the noisy channel conditions encountered in real-world deployments.

The encoder portion of a trained autoencoder captures a generic, device-agnostic representation of nonlinear distortion. These pretrained weights can be frozen or used as a warm start when training a predistorter for a different power amplifier from the same family, dramatically reducing the volume of new labeled training data and the time required for model extraction on the target device.

After unsupervised pretraining, the decoder is discarded and replaced with a predistortion output layer. The entire network is then fine-tuned using a small set of labeled data in a Direct Learning Architecture (DLA). This semi-supervised approach combines the data efficiency of unsupervised learning with the precision of supervised error minimization, yielding high ACLR improvement with limited labeled captures.

The reconstruction loss used during pretraining acts as an implicit regularizer, guiding the network toward a parameter space that represents the true data-generating distribution of the PA rather than overfitting to a specific predistortion target. This geometric constraint on the weight manifold leads to smoother predistorter transfer functions and improved generalization across varying signal statistics.

A methodology where a neural network predistorter trained on one power amplifier is partially reused as a starting point for training on a different PA. The encoder's compressed representation of nonlinear behavior, learned during unsupervised pretraining, serves as a highly transferable feature extractor. This significantly reduces the data volume and training time required for model extraction on new amplifier hardware.

A modeling failure where the neural network predistorter memorizes the training data's noise and specific signal characteristics rather than learning the true underlying PA nonlinearity. Autoencoder pretraining acts as a powerful regularizer by forcing the network to learn a compressed, noise-agnostic latent representation of the amplifier's behavior, inherently discarding irrelevant signal-specific artifacts before fine-tuning for the predistortion task.

A deep neural network design where layers learn the difference between the target and the input, implemented via skip connections. In autoencoder linearization, residual blocks within the encoder-decoder structure simplify the optimization landscape by allowing the network to learn the deviation from a linear pass-through, making it easier to model the subtle nonlinear distortions introduced by the power amplifier.

The artificial expansion of a PA measurement dataset by applying transformations like phase rotation, amplitude scaling, and additive noise injection. When training an autoencoder for behavioral modeling, data augmentation forces the bottleneck layer to learn invariant features of the amplifier's distortion characteristic, improving the robustness and generalization of the final predistorter to unseen signal conditions.

The process of reducing the numerical precision of a neural network's weights and activations from 32-bit floating-point to lower bit-width integers (e.g., INT8). For autoencoder-based predistorters deployed on FPGAs, quantization-aware training ensures the compressed latent representation remains discriminative even under severe precision constraints, directly decreasing inference latency and memory footprint.

An adaptive training paradigm where the neural network predistorter coefficients are continuously updated during live signal transmission. The autoencoder's pretrained decoder provides a stable, physically consistent baseline model of the PA's forward behavior. Online adaptation can then focus on fine-tuning only the predistorter weights to track time-varying characteristics due to temperature drift and device aging, preventing catastrophic forgetting.