Cross-Layer Equalization

CLE exploits the scale-equivariance property of activation functions like ReLU. By scaling up weights in one layer and inversely scaling them down in the next, the mathematical output remains identical while the weight distributions are balanced. This directly addresses the problem where one channel has a large dynamic range, forcing the entire tensor's quantization grid to accommodate outliers at the expense of precision for smaller values.

During the equalization process, bias terms must be adjusted to maintain functional equivalence. The bias in the subsequent layer is modified to absorb the scaling factor applied to the preceding layer's weights. This ensures that the equalized model produces bitwise-identical outputs to the original full-precision model before quantization is applied, preserving the network's learned decision boundaries.

CLE is particularly effective after Batch Normalization Folding. When BN parameters are absorbed into the preceding convolutional layer, large folded bias values can emerge. These high biases create extreme activation ranges. CLE redistributes these ranges across layers, preventing the folded biases from dominating the quantization grid and causing severe accuracy collapse in low-precision integer formats.

Unlike Quantization-Aware Training (QAT) or calibration-based Post-Training Quantization (PTQ), CLE requires no representative dataset. It operates purely on the model's weight tensors using analytical methods. This makes it ideal for deployment scenarios where access to the original training or calibration data is restricted due to privacy, security, or bandwidth constraints on the target FPGA platform.

While CLE was designed to enable effective per-tensor quantization, it also enhances per-channel weight quantization. By equalizing the weight ranges, CLE reduces the variance in scaling factors across channels. This leads to a more uniform utilization of the integer representation range, further minimizing the signal-to-quantization-noise ratio (SQNR) degradation in the deployed integer-only inference pipeline.

A specific variant of CLE targets layers with extreme bias values after Batch Normalization folding. The technique identifies layers where the bias term is significantly larger than the weight variance. It then applies a targeted scaling transformation to absorb this high bias into the subsequent layer's weights, effectively neutralizing a primary source of activation range outliers without altering the network's mathematical function.

A pre-deployment optimization that mathematically absorbs the parameters of a batch normalization layer into the preceding convolutional layer's weights. During inference, batch normalization applies a fixed linear transformation: y = γ(x - μ)/σ + β. By folding these scale (γ) and shift (β) parameters into the weight and bias terms of the preceding layer, the normalization computation is eliminated entirely at runtime. This reduces memory access and compute overhead, and is a critical prerequisite for cross-layer equalization, as the equalization scaling factors must be absorbed into the folded weights to maintain mathematical equivalence.

A training methodology that simulates the effects of low-precision arithmetic during both the forward and backward passes. Unlike post-training quantization, QAT inserts fake quantization nodes that model rounding and clipping errors, allowing the optimizer to adapt the weight distribution to minimize accuracy loss. For RF modulation classifiers, QAT typically preserves <0.5% accuracy degradation at INT8 precision. Cross-layer equalization can be applied as a preprocessing step before QAT to reduce the dynamic range disparity that fake quantization nodes must compensate for, accelerating convergence.

Per-tensor quantization assigns a single scale factor and zero-point to an entire activation tensor, while per-channel quantization assigns independent scale factors to each output channel. Per-tensor is hardware-efficient but suffers severe accuracy degradation when activation ranges vary significantly across channels—a common scenario in depthwise separable convolutions used for RF inference. Cross-layer equalization directly addresses this limitation by redistributing the dynamic range across consecutive layers, making per-tensor quantization viable without the hardware complexity of per-channel scaling.

An automated design flow that translates algorithmic descriptions written in C or C++ into register-transfer level (RTL) hardware descriptions for FPGA implementation. Tools like Vitis HLS and hls4ml enable signal processing engineers to deploy quantized neural networks without manual Verilog or VHDL coding. Cross-layer equalization simplifies the HLS pipeline by ensuring that activation tensors fit within uniform fixed-point representations, avoiding the need for custom data path widths per layer that complicate synthesis and reduce achievable clock frequency.

A factorized convolutional operation that splits standard convolution into two stages: a depthwise spatial filter (one kernel per input channel) and a pointwise projection (1×1 convolution across channels). This reduces parameter count by approximately 8-9× compared to standard convolutions, making it a cornerstone of mobile and edge RF classifiers. However, the depthwise stage produces highly imbalanced output channel ranges, making cross-layer equalization particularly effective when applied between the depthwise and pointwise layers to equalize the per-channel activation distributions before quantization.

A deployment mode where all neural network arithmetic—matrix multiplications, activations, and element-wise operations—is performed using integer operations exclusively. This eliminates the need for floating-point units (FPUs) and enables efficient execution on FPGA DSP slices configured as fixed-point multipliers. Cross-layer equalization is a key enabler for integer-only inference because it minimizes the quantization error that accumulates when floating-point dynamic ranges are collapsed into uniform INT8 representations, preserving the signal-to-quantization-noise ratio (SQNR) required for accurate modulation classification.