ASDL trims CNN coefficients to about 1.5 active bits: ImageNet accuracy slips from 81.3% to 80.41% with no retraining
Synopsis
Multiply-accumulates dominate convolutional-network inference, and each one can be rewritten as a handful of shift-and-adds whose count equals the number of signed powers of two (SPT) in the coefficient. ASDL replaces the earlier greedy routine that inserted one SPT per step with a routine that fills an entire bit plane per step, converting faster than MPGBP on ResNet50's longest coefficient vector while keeping a lower approximation error. Applied to pre-trained CNN coefficients, it leaves about 1.5 active bits per coefficient on average and moves ImageNet accuracy from 81.3% to 80.41% without retraining or calibration.
Interpretation
A convolutional network's multiply-accumulates can be rewritten as shift-and-adds, and the number of shifts equals the count of signed powers of two in the multiplier coefficient, so lowering that count directly lowers the arithmetic. Earlier routines inserted one signed power of two at a time, searching the residual vector for its largest magnitude on every pass, so the number of passes grew with the target count. On an NVIDIA A100, timed over 100 repeated conversions of ResNet50's longest coefficient vector, 512 vectors of 4,608 parameters each in bfloat16, ASDL converted faster than MPGBP.
Instead of placing a single power of two per pass, ASDL writes every position sharing one exponent at once, filling one bit plane per pass. That turns each pass from a repeated single-point maximum search into grouping by exponent followed by one partial sort of candidate positions, keeping only the largest magnitudes when the budget would be exceeded. The algorithm runs in , and dropping the candidate-reduction step to allow under-allocation brings it to $O(n)$.
Applied to pre-trained CNN coefficients, ASDL moves ImageNet accuracy from 81.3% to 80.41% at roughly 1.5 active bits per coefficient on average. That accuracy comes with no retraining or calibration, the two steps the most aggressive quantization methods typically require. Tests cover public pre-trained ResNet, ShuffleNet and EfficientNet architectures, with coefficients and activations in bfloat16 and approximation applied to coefficients but not biases.
Perspective
The demonstrated setting is post-training coefficient approximation for ImageNet image classification across public pre-trained ResNet, ShuffleNet and EfficientNet models, with coefficients and activations in bfloat16 and approximation applied to coefficients but not biases. The authors list testing the method on other network architectures and applications as a future direction.
The 3.33x complexity reduction is an estimate made under assumptions including shift-and-add multiplication and 24-bit original coefficient precision, not a measured hardware result. EfficientNet's squeeze-excite and SiLU involve multiplications between dynamic values and cannot be fully converted to multiplierless designs under the current ahead-of-time vector approximation. The authors also note that larger models appear more resilient to fewer active bits: at the same setting ShuffleNet produces incoherent results while EfficientNet holds about 70% accuracy.
