A preprint reports a sharp accuracy cliff in a new approach to neural-network compression. Across the three networks tested, accuracy stayed nearly unchanged at small values of the method’s error allowance, then fell dramatically at a critical value between 0.35 and 0.5. The result comes from post-training compression trials on three trained networks, not from a broad survey of models.
The method treats a network as a graph whose links carry labels that can be combined. In the paper’s terminology, an M-fibration requires the label on a target link to match the combined labels of the incoming links lifted to each node that maps to it. That extends graph fibrations to weighted or algebraically labelled graphs, while the approximate version permits a measured mismatch.
A formal error budget
On the exact side of the framework, the paper proves that every finite M-graph has a minimum base and a minimal M-fibration. The base is unique up to isomorphism, and the fibration is unique up to composition with an isomorphism of that base. These are mathematical guarantees under the paper’s definitions and finite-graph assumptions, rather than estimates from the network tests.
To handle approximate compression, the paper defines a local error at each node and a global error for the full mapping. It then defines an epsilon-approximate M-fibration as one whose global error stays within the chosen threshold, along with a fibration distance between graphs. A function called beta G of epsilon records the smallest number of target-graph nodes attainable within that error budget. In practical terms, the framework makes the trade-off between allowed mismatch and representation size explicit.
The partitioning procedure is greedy. It groups nodes using local in-weight vectors, a chosen center and an epsilon-radius rule, and guarantees that the resulting radius is no greater than epsilon. But the paper says the procedure may not produce the partition with the fewest possible classes. The formal error guarantee and the most compact possible partition are therefore not the same promise.
The compression results
The empirical demonstration covered two MNIST networks and one CIFAR-10 network. Each was compressed after training, converted back to a neural network and evaluated on its test set. The comparison used the corresponding uncompressed versions as the baseline.
In the MLP, an error allowance of 0.4 left 76% of units and 79% of parameters, with reported accuracy of 97.53%, compared with 97.76% before compression. At 0.5, the retained share fell to 42% of units and 45% of parameters, while accuracy was 79.6%.
The CNN also showed a large gap between a moderate and more aggressive setting. At 0.4, it retained 94% of units and 93% of parameters and recorded 98.89% accuracy, compared with 99.09% uncompressed. At 0.5, only 10% of units and 2.6% of parameters remained, and reported accuracy was 14.7%.
For VGG16-BN, the reported settings were 0.30, 0.35 and 0.40. The network retained 69%, 62% and 48% of units, respectively, and 55%, 44% and 28% of parameters. Accuracy was 91.53%, 90.18% and 10.0%, against 94.16% uncompressed.
The authors describe this pattern as a phase transition: a stretch of near-stable accuracy followed by a steep drop. They say the sharp edge reflects the certified worst-case criterion used by the method. They also report that heuristic clustering based on mean or least-squares representatives produces a smoother accuracy-versus-size decline, but without the same guarantee.
What the evidence does not settle
The evidence remains narrow. It covers three trained networks, and the experiment was conducted after training. The paper explicitly says this setup is not comparable to the method described in [2], so these figures should not be read as a general test of training-time compression.
The paper says it provides a self-contained Python library implementing the algorithms and an ONNX script that accepts networks, compresses them, tests accuracy and reports the results. That makes the claimed workflow available for inspection, but the supplied evidence does not include independent replication.
For now, the study’s contribution is a formal language for controlled graph reduction, together with a demonstration of how quickly accuracy can fall when compression passes a model-specific boundary. The results define an explicit error budget, but they do not establish that the method is optimal for every error setting.
Paper data and sources
Original title: M-Fibration Theory with Applications to Neural Network Compression
Authors: Paolo Boldi
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text