Original title: Improving Accuracy and Efficiency in DNNs with Approximate Multipliers: Insights from Information Bottleneck Theory
Salar Shakibhamedan, Nima Amirafshar, Axel Jantsch, Nima TaheriNejad · September 2026
Deep neural networks use massive amounts of energy to multiply numbers. Engineers can use approximate multipliersCircuits that perform multiplication quickly and with less power by deliberately rounding the result slightly., which are shortcut math circuits that save power by making tiny computational errors. The problem is that these errors usually hurt the network's accuracy. In this paper, researchers found a way to fix that. Instead of treating every layer the same, they applied varying levels of approximation across different network layers using 8-bit quantized numbers. This smart mix actually improved the network's feature extraction and information flow. As a result, the models cut energy use by up to 44.75% across evaluated setups while boosting Top-1 accuracy by up to 2.07 percentage points over exact baselines, improving both efficiency and performance together.
Imagine packing a suitcase. If you pack every item exactly as it is, the suitcase becomes too heavy. But you can slightly compress some less important items without losing their function – like rolling towels instead of folding them – while leaving important items like a jacket uncompressed. The suitcase becomes lighter and you still have everything you need. Here, the suitcase is a neural network, the items are calculations in different layers, and the compression represents computing with small errors. By applying different levels of error to different layers, the network saves energy and sometimes becomes more accurate.
Deep neural networks power many applications but consume lots of energy. Approximate multipliers can save energy but usually hurt accuracy. This work shows the opposite is possible: by carefully choosing which layers to approximate, both energy and accuracy improve, making efficient AI more practical.
The abstract only reports results on VGG models and uses estimated energy savings. Accuracy improvement is modest (up to 2.07 percentage points) and may not generalize. The Information Bottleneck analysis is qualitative. The MAC-count comparison is for computational workload only, not a direct speedup measure.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Improving Accuracy and Efficiency in DNNs with Approximate Multipliers: Insights from Information Bottleneck Theory — Salar Shakibhamedan, Nima Amirafshar, Axel Jantsch et al. — AI — doi:10.3390/ai7090385
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.