Monash Researchers Show Pruned Hydra Beats Accurate Models on Energy
Monash and German researchers prune the Hydra and Quant time series classifiers, cutting inference energy by up to 70% with only a small loss in accuracy, and show that hardware choice can change which model looks best.
Takeaways
- Researchers prune Hydra and Quant, cutting time series inference energy by up to 70%.
- At 80% pruning, models keep 92 to 97% of quality while using 50 to 70% fewer resources.
- The pruning ranks feature blocks with a ridge model, with a distribution-free error bound in the linear case.
- Hydrant stacks Hydra class scores onto Quant features, improving accuracy but costing significantly more energy.
- More than 4,000 runs on 20 MONSTER datasets show that CPU, GPU and batch size choices change rankings.
- The code, logs and STREP dashboard are open-source, with MONSTER data on Hugging Face.
Time series classification labels sequences of sensor readings, such as assigning sleep stages to EEG data or land-cover types to satellite histories. Researchers usually compare classifiers by accuracy and total runtime on one machine. A paper from Monash University, TU Dortmund, Johannes Gutenberg University Mainz, and University Medicine Essen adds inference energy, memory, hardware, and batch size to the evaluation. Under that broader scorecard, pruned versions of Hydra and Quant outperform several more accurate but resource-intensive models.
The study evaluates more than 4,000 configurations, introduces a pruning method with a provable bound for linear models, and presents Hydrant, a stacked combination of Hydra and Quant. At an 80% pruning rate, the tested models retained 92% to 97% of their original quality on average while reducing measured resource use by 50% to 70%.
Small benchmarks hide deployment cost
Most time series classification research emphasizes accuracy and total experimental runtime. Energy is rarely reported, hardware comparisons are uncommon, and training cost is often combined with inference cost. That accounting obscures production workloads, where a model may process enough samples for inference energy to exceed its one-time training cost.
The classic UCR and UEA archives have median dataset sizes of 217 and 255 examples, respectively, leaving little room to study sustained inference. This paper uses the larger MONSTER archive and averages results across its five predefined data splits.
| Dataset property | Range |
|---|---|
| Instances | 9,000 to 200,000 |
| Classes | 2 to 82 |
| Channels | 1 to 113 |
| Series length | 23 to 5,000 time steps |
| Datasets | 20 |
Redundancy creates room to prune
Hydra and Quant transform each time series into many engineered features, then train a comparatively simple classifier. Hydra applies groups of random convolution kernels and counts how often each kernel produces the strongest or weakest response within its group. A ridge classifier predicts from those counts. Quant calculates quantile statistics over many intervals of the original and transformed series, then feeds them to randomized decision trees.
Both methods organize features into natural blocks: kernel groups for Hydra and intervals for Quant. The paper tests whether blocks with little influence on an intermediate linear classifier can be removed. Every discarded block eliminates feature computation for each subsequent inference.
A bound on linear score drift
The pruning pipeline ranks feature blocks by their fitted coefficients:
- Compute the complete feature set and fit an intermediate ridge classifier.
- Score each block using the mean absolute value of its coefficients.
- Keep the highest-scoring (1 − ζ) share of blocks, where ζ is the pruning rate.