Automated bird-migration monitors listen for months and mostly hear nothing. In the BirdVox recordings, flight calls fill less than 0.1% of the audio. Yet the standard pipeline runs a full detector on every frame and a full species classifier on every detection. This proposal asks for a classifier whose computation follows the biological label tree. The first question the tree asks, "is there a call at all?", is answered by the first few layers of the network. A second exit, after a few more layers, answers "which bird family?". The small, specialised layers that separate sister species run only when the network is already confident about the family. When the device cannot afford, or cannot trust, a species answer, it reports the family or order instead of a wrong species. Two pieces make this plausible. The first is per-channel energy normalization (PCEN), an audio front end that divides each frequency band by a running estimate of its own background level, so steady noise such as wind or traffic flattens out and brief calls stand out. PCEN is exactly what should let a very shallow first exit reject noise reliably. The second is hierarchical selective classification, which already shows how to back off to a coarser, safer label under uncertainty.
The claim is that aligning early exits with taxonomic ranks gives a better trade-off than two simpler options: compressing a flat model and backing off afterwards, or using generic early exits that predict every label at every depth. The quantity being traded is how specific the correct answers are at a guaranteed error rate, per unit of on-device compute. The second claim is that when compression hurts, the damage appears as retreats to coarser ranks rather than as confident wrong species. That matters because the rare species, the ones with the fewest training examples, are the ones pruning tends to destroy.
What exists and what is missing
The acoustic side is well supplied. Long-distance Detection of Bioacoustic Events with Per-channel Energy Normalization shows that PCEN cuts false alarms 50x in marine near-field noise and 5x for nighthawks at 500 m compared with a plain logarithm. That work covers only binary detection, with no classifier and no compression. Trainable Frontend For Robust and Far-Field Keyword Spotting learns PCEN parameters jointly with the network, cutting false rejections by 14% absolute, and notes that frozen PCEN adds no parameters. It studies a single keyword and never compresses the backbone. The closest system is BirdVoxDetect and BirdVoxClassify (BirdVoxDetect; BirdVox). It uses PCEN, reaches 52% detection F1 on BirdVox-296h, and raises species F1 on BirdVox-14SD-v1.1 from 66.71% to 72.82% by applying taxonomic consistency rules after prediction. That system is our starting point and our main reference number. It is a dense two-model pipeline, the hierarchy lives in post-hoc rules rather than in the computation, and it reports no compression.
On the hierarchy side, three papers supply methods we reuse. Coherent Hierarchical Multi-Label Classification Networks forces a parent's probability to be at least as large as its most likely child's, with no added parameters, using its Max Constraint Module. It was evaluated on text, tabular and medical-image data with dense backbones. Hierarchical Selective Classification climbs from the most probable leaf toward the root until enough probability mass accumulates. Its thresholds are calibrated with split conformal prediction, a held-out-data procedure that guarantees an error rate. It gains 14.94% in hierarchical coverage over flat abstention on ImageNet, but it treats the backbone as a frozen black box that always runs to completion. HFT-CNN supplies the tail-class remedy: initialise child classifiers from their data-rich parents (Macro-F1 51.40 vs 47.71 flat on RCV1). Its finding that hard parent gating hurts recall constrains our routing design.
On the compute side, Model Cascading and Distilled Hierarchical Neural Ensembles save compute by escalating uncertain inputs to more capacity. Model Cascading matches BERT-Large on QQP at 11.07% of its computation. Their trees are computational graphs, and every exit predicts the full flat label set. This is the sharpest contrast with our proposal. A generic cascade escalates when the model is unsure. A taxonomic exit escalates when the model is sure of the family and still needs the species, and it retreats upward when even the family is unsure. Which escalation direction is right for fine-grained, long-tailed audio is an open empirical question, and we test both. Finally, RestoreLCC and Super Tickets evaluate compression on aggregate accuracy only. Nobody in this literature has measured how compression damage spreads across taxonomic ranks and class frequencies.
Core Claim
A single network built as a shared trunk, with exits at the detection, order/family and genus/species ranks and small per-family branches, dominates three alternatives at matched average compute including the front end. The alternatives are a compressed flat model with post-hoc climbing, a small dense model with climbing, and a generic early-exit model with climbing. Dominance is measured as specificity-at-guaranteed-risk per multiply-accumulate. We further predict that PCEN, compared with log-Mel, specifically raises the fraction of noise frames the shallowest exit can reject at fixed call recall. If so, the front end and the exit structure reinforce each other rather than simply adding.
Research Program
The model has four parts. The front end is trainable PCEN learned jointly with the network, as in the keyword-spotting work. The trunk is a few convolutional blocks. After the first block, a tiny binary head answers call versus no-call; this is the root of a hierarchical binary decomposition. After the second block, a head predicts order and family. Each family owns a small branch for genus and species. That branch is initialised from the family head's features, following the HFT-CNN idea, so rare species borrow statistical strength. All heads are made coherent with the Max Constraint Module. Hard top-down gating would propagate family errors irreversibly, so a family's branch runs whenever that family's probability exceeds a routing threshold. This allows more than one branch to run on ambiguous clips.
At inference, the conformal climbing rule is applied at each exit as it is reached, which turns it into a compute policy. We compare two variants. Under the refine-when-confident policy, computation continues only if the current rank clears its calibrated threshold, and otherwise the node's parent is emitted. Under the escalate-when-unsure policy, computation continues when the current rank is uncertain, in the style of generic cascades.
Compression is applied deliberately and asymmetrically. The trunk and the early heads see all data and carry most of the traffic, so they are pruned and quantised to INT8. The per-family branches are small and rarely run, so they stay at higher precision. PCEN itself is implemented in fixed point, with the IIR smoother and the power-law compression approximated in integer arithmetic. Its cost is counted in the budget, which answers the open question of whether a quantised front end becomes the bottleneck or harms the representations downstream.
Before building anything, a diagnostic study measures what uniform compression does to the existing dense BirdVox-style classifier. We apply magnitude pruning at several sparsities and INT8/INT4 quantisation, then report F1 per taxonomic rank and per species binned by training-set frequency. This tests the premise that compression concentrates damage on data-sparse leaves while sparing families. If the premise fails and damage is uniform, the rationale for rank-asymmetric compression weakens, and we would report that.
Baselines
The naive reference is the dense BirdVoxDetect-plus-BirdVoxClassify pipeline with post-hoc taxonomic rules, reproduced on the same splits with the same PCEN front end. Its reported numbers are 52% detection F1 on BirdVox-296h (BirdVoxDetect) and 72.82% species F1 on BirdVox-14SD-v1.1 with consistency rules (BirdVox). We found no reported result on Lacuna for the compute cost or the compressed accuracy of either BirdVox model, so we measure both ourselves. Beating this reference means matching 72.82% species F1 and 52% detection F1 at a clearly lower average cost per hour of audio, for example under a third, or exceeding both at equal cost.
The strongest competitor, in our view, is that same dense pipeline compressed to our target budget by pruning and INT8 quantisation and then wrapped in the Hierarchical Selective Classification climbing rule with conformal calibration. It combines the best documented backbone, standard compression and the best documented back-off method, and differs from our proposal only in not tying computation to rank. It receives the same training data, the same number of tuning trials for sparsity and quantisation settings, and the same calibration split. A win means higher hierarchical coverage at the same guaranteed risk (target 5% error) at equal or lower measured compute. We also require species F1 on the rarest frequency bin to be at least as good.
Two further competitors address specific parts of the claim. A small dense model trained from scratch at the same compute, with climbing, tests whether compression is needed at all. A generic early-exit model with the same trunk, whose exits all predict the full label set, is gated by a distance-to-uniform confidence score as in Model Cascading, and uses climbing on its final output. This last model is the ablation that isolates our key ingredient, taxonomic exits versus merely early exits.
The remaining ablations each switch off one component. Log-Mel replaces PCEN. The Max Constraint Module is removed. Parent-inherited initialisation of leaf branches is removed. Uniform compression replaces rank-asymmetric compression. Floating-point PCEN replaces fixed-point PCEN. The two escalation policies are compared head to head.
Why not just…?
Why not just compress the existing model and apply hierarchical selective classification afterwards? This is the most serious threat, and it is our main baseline rather than a strawman. The evidence against it is indirect. Both the compression and hierarchy literatures suggest that uniform pruning disproportionately harms fine, data-sparse boundaries, so climbing would mostly recover coarse answers the model spends full compute to produce. The post-hoc approach also saves nothing on the 99.9% of frames that contain no call. The diagnostic study and the matched-compute comparison settle the question directly. If climbing on a compressed flat model matches taxonomic exits in coverage per compute, the architectural claim is wrong, and we would report the diagnostic study as the contribution.
Why not just train a small model from scratch? Distilled ensembles beat compact from-scratch designs on ImageNet by 1.5% to 11% at equal compute (Distilled Hierarchical Neural Ensembles). That evidence comes from balanced vision data, so it does not transfer automatically. A small model spends the same compute on every frame, and its benefit on 14 species could be larger than on ImageNet. The small-dense baseline with climbing, run at three budgets, measures this.
Why not just use a generic cascade that escalates when uncertain? Model Cascading saves 88.93% of compute on QQP with flat labels. Under a taxonomy, though, uncertainty about the family may mean the clip is too faint or distant for any depth to resolve. In that case escalation wastes compute, and the right move is to answer at a coarser rank. The generic early-exit baseline, and the comparison between the two escalation policies, test this. If escalate-when-unsure wins, we adopt it inside the taxonomic structure and the claim narrows to "rank-aligned exits" alone.
Why not just use log-Mel with data augmentation? On BirdVox-296h, traditional features reach 3% F1 versus 52% with PCEN (BirdVoxDetect). Our specific question is whether PCEN makes the shallowest exit trustworthy, which the root-exit-rate check below measures.
Proposed Experiments
Training uses BirdVox-full-night and BirdVox-222k. Detection is evaluated on BirdVox-296h and classification on BirdVox-14SD-v1.1 with its Order→Family→Genus→Species taxonomy. As a robustness check, CONI-Knight at 30 to 500 m tests whether the root exit degrades gracefully with distance. BirdVox-14SD has a shallow taxonomy and few species, which limits the claim. The results should be read as evidence for long-tailed monitoring vocabularies, not as proof at iNaturalist scale.
The metrics are:
- detection F1;
- per-rank and frequency-binned species F1;
- the hierarchical risk-coverage curve and coverage at 5% guaranteed error;
- the fraction of errors that are wrong leaves versus retreats;
- measured multiply-accumulates and latency per hour of audio, including PCEN;
- peak memory, since cascades inflate storage.
Three results would decide the proposal:
- At matched compute, taxonomic exits must beat compressed-flat-plus-climbing in coverage at 5% risk, without losing rare-bin species F1.
- Under increasing compression, the taxonomic model's errors must shift toward retreats while the flat model's shift toward wrong species.
- PCEN must let the first exit reject a markedly larger share of noise frames than log-Mel at the recall of the dense reference detector.
If the first result fails, the architecture is not worth its complexity. If only the third holds, the finding is narrower: a PCEN-gated cheap detector, which is still useful for field sensors.
References
- Long-distance Detection of Bioacoustic Events with Per-channel Energy Normalization
- Trainable frontend for robust and far-field keyword spotting
- BirdVoxDetect: Large-Scale Detection and Classification of Flight Calls for Bird Migration Monitoring
- BirdVox: Machine listening for bird migration monitoring
- Coherent Hierarchical Multi-Label Classification Networks
- Hierarchical Selective Classification
- HFT-CNN: Learning Hierarchical Category Structure for Multi-label Short Text Categorization
- Model Cascading: Towards Jointly Improving Efficiency and Accuracy of NLP Systems
- Distilled Hierarchical Neural Ensembles with Adaptive Inference Cost
- Restoring Pruned Large Language Models via Lost Component Compensation
- Super Tickets in Pre-Trained Language Models: From Model Compression to Improving Generalization