Rethinking the Role of the Classification Head in CNN-BiLSTM-Based Human Activity Recognition
Year: 2026
Type: article
Abstract: Wearable sensor-based human activity recognition (HAR) underpins healthcare monitoring, fall detection, and fitness tracking. HAR pipelines pair a feature-extraction backbone with a classification head. Backbone design is well studied, whereas the classification head has received less attention. We ask how much the choice of the classification head matters. Holding the CNN-BiLSTM backbone architecture fixed across seven public HAR datasets under leave-one-subject-out cross-validation, we compare four heads: a 774-parameter linear projection, a small multi-layer perceptron (MLP), a Matched MLP, and a parameter-matched Gaussian radial-basis-function Kolmogorov–Arnold network (KAN). The classification head has a limited effect. The best-to-worst spread across heads is under 3 percentage points (pp), and the 774-parameter linear head stays within 1.5 pp of the best head on every dataset. When the backbone is frozen and each head is retrained on the fixed features, the spread narrows to about 1 pp on six of seven datasets, indicating that additional head capacity has little effect once the representation is fixed. Under parameter matching, the KAN head ranks below the Matched MLP on six of seven datasets but first on HHAR, with effect sizes near zero in every KAN–MLP comparison. A minimal linear head therefore matches the accuracy of more complex heads at lower full-model latency than the KAN head on both GPU and CPU (2.10 versus 3.90 ms and 4.37 versus 5.55 ms), and at 0.5% of the KAN head’s parameter count.
Source: Electronics
Authors: Junbing Shi, Yingmin Wang, Libo Yang
Institution: Taiyuan University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7225652108