🤖 machine learning

Dywave: Event-Aligned Dynamic Tokenization for Heterogeneous IoT Sensing Signal

Dywave is a novel dynamic tokenization framework for heterogeneous IoT sensing signals that leverages wavelet-based hierarchical decomposition to align input representations with intrinsic temporal events, achieving significant improvements in accuracy and computational efficiency compared to state-of-the-art methods.

Original authors: Tomoyoshi Kimura, Denizhan Kara, Jinyang Li, Hongjue Zhao, Yigong Hu, Yizhuo Chen, Xiaomin Ouyang, Shengzhong Liu, Tarek Abdelzaher

Published 2026-05-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tomoyoshi Kimura, Denizhan Kara, Jinyang Li, Hongjue Zhao, Yigong Hu, Yizhuo Chen, Xiaomin Ouyang, Shengzhong Liu, Tarek Abdelzaher

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Uniform Slice" Mistake

Imagine you are trying to summarize a 30-minute movie for a friend.

The Old Way (Current Methods): You cut the movie into 30 equal 1-minute chunks, regardless of what is happening.
- The Problem: If minute 5 is a slow, quiet scene where a character just sits and thinks, and minute 15 is a massive explosion with car chases, your old method treats them exactly the same. You spend the same amount of time describing the boring sitting as you do the explosion. This wastes your time and confuses your friend because the "important" parts get lost in a sea of "boring" details.

In the world of Internet of Things (IoT) sensors (like smartwatches or environmental monitors), computers currently do this exact thing. They chop up continuous streams of data (like heartbeats or movement) into fixed-size, equal-length pieces. But real life isn't uniform. A person might walk for 10 seconds, then stop for 5 minutes, then wave their hand for 1 second. The old method forces the computer to process the 5 minutes of "doing nothing" just as heavily as the 1 second of "waving."

The Solution: Dywave (The "Smart Editor")

The authors propose Dywave, a new way to process these signals. Instead of cutting the data into equal slices, Dywave acts like a smart movie editor that only cuts when something interesting happens.

Here is how it works, step-by-step:

1. Listening to the "Layers" of Sound (Hierarchical Embedding)

Imagine you are listening to a symphony. You can hear the deep bass (the slow, long trends) and the high-pitched violins (the quick, sharp notes).

What Dywave does: It uses a mathematical tool called a "Wavelet Transform" to separate the sensor data into these different layers. It looks at the slow, steady background (like a person standing still) and the fast, sharp spikes (like a sudden jump) separately. This helps the computer understand the "shape" of the event better than just looking at the raw numbers.

2. Finding the "Punchlines" (Temporal Anchor Formation)

Now, imagine you are watching that movie again. You want to find the moments where the story actually changes.

What Dywave does: It scans the data to find "anchors." These are the exact moments where the signal changes significantly—like the moment a person starts walking or the moment a door slams.
The Magic: It ignores the boring parts. If a person is just standing still for 10 seconds, Dywave realizes, "Nothing new is happening here," and decides not to make a new "cut" or "token" for every single second of that silence. It only marks the transitions.

3. The "Highlight Reel" (Dynamic Temporal Fusion)

Finally, Dywave creates a summary.

What Dywave does: It groups the boring, repetitive seconds together into a single, compact "token" (a digital word). But for the exciting, fast-changing moments, it keeps the details sharp.
The Result: Instead of feeding the computer 1,000 data points (where 900 are just "standing still"), Dywave feeds it 200 points. But those 200 points are packed with the most important information. It's like turning a 3-hour movie into a tight 20-minute highlight reel that still tells the whole story perfectly.

Why This Matters (The Results)

The paper tested Dywave on real-world data, such as:

Human Activity: Recognizing if someone is cleaning, reading, or walking.
Stress Monitoring: Detecting stress levels from heart signals.
Object Detection: Sensing moving objects nearby.

The findings were impressive:

Smarter: Dywave was up to 12% more accurate than the best existing methods. Because it didn't get confused by "boring" data, it could focus on the real events.
Faster & Cheaper: It reduced the amount of data the computer had to process by up to 75%. This means devices can run longer on batteries and respond faster because they aren't wasting energy on empty space.
Robust: It worked well even when the data was noisy or when the length of the recording changed.

The Bottom Line

Think of Dywave as a smart filter for the internet of things. Instead of blindly chopping up sensor data into equal, rigid blocks, it listens to the data, finds the "story," and creates a custom summary that highlights the action and skips the silence. This makes AI systems smarter, faster, and more efficient at understanding the physical world.

Technical Summary: Dywave: Event-Aligned Dynamic Tokenization for Heterogeneous IoT Sensing Signals

1. Problem Statement

Internet of Things (IoT) systems generate continuous streams of heterogeneous sensing signals (e.g., IMU, ECG, acoustic) to support applications like human activity recognition (HAR) and environmental monitoring. Unlike natural language (words) or vision (spatial patches), these signals lack intuitive, discrete semantic units. They are inherently non-stationary, multi-scale, and temporally heterogeneous, with information encoded in transitions between physical events rather than uniform intervals.

Current state-of-the-art approaches rely on uniform windowing (fixed-size patching) to tokenize these signals. This method presents three critical limitations:

Semantic Misalignment: Fixed boundaries rarely align with intrinsic physical event transitions, leading to event fragmentation and the obscuring of underlying semantics.
Inefficiency: Uniform patching allocates equal computational resources to redundant, quiescent intervals and salient, dynamic bursts, inflating input token lengths and computational costs without adding information.
Hyperparameter Sensitivity: Optimal patch sizes and strides are application-specific and require extensive, time-consuming per-domain tuning, with no universally effective configuration.

This creates a "tokenization gap" where raw sensing signals cannot be effectively processed by large-scale sequence models without losing physical coherence or incurring excessive latency.

2. Methodology: Dywave

The authors propose Dywave, a dynamic tokenization framework that constructs compact input representations aligned with intrinsic temporal structures and underlying physical events. Dywave operates as a modular front-end compatible with mainstream backbone encoders (e.g., Transformers, Mamba2) without architectural changes. The pipeline consists of four stages:

2.1. Hierarchical Embedding Extraction

Dywave leverages the Maximal Overlap Discrete Wavelet Transform (MODWT) to decompose raw signals into multi-resolution time-frequency representations. Unlike standard DWT, MODWT is undecimated, preserving the original sequence length.

Detail Stream: Captures high-frequency transients using lightweight convolutional layers.
Context Stream: Encodes slow-varying, long-range trends using a lightweight hourglass transformer with adaptive downsampling.
Fusion: These streams are concatenated to form a unified hierarchical embedding ( $E_F$ ) that captures both fine-grained transients and global context.

2.2. Temporal Anchor Formation

Instead of fixed segmentation, Dywave adaptively identifies temporal anchors at semantic transitions.

Saliency Estimation: The model estimates event saliency ( $P_t$ ) by measuring the representational change between adjacent timesteps using cosine similarity between projected key and query vectors of the hierarchical embeddings. High saliency indicates abrupt representational shifts (events).
Anchor Selection: To avoid noise and over-allocation, Dywave applies temporal Non-Maximum Suppression (NMS) followed by Top-K selection to extract a compact set of anchors ( $A$ ). The number of anchors is bounded by a compression ratio $\tau$ .

2.3. Dynamic Temporal Information Fusion

Once anchors are identified, Dywave aggregates neighboring timesteps into anchor-aligned tokens.

Cluster Formation: Each timestep is assigned to its nearest anchor based on temporal proximity.
Saliency-Weighted Pooling: Embeddings within each cluster are aggregated using saliency scores as weights. This prioritizes informative transitions while compressing stable, redundant intervals.
Output: The result is a compact sequence of tokens ( $E$ ) whose length adapts to semantic complexity rather than raw signal duration.

2.4. Training Objective

To ensure the compressed tokens retain multi-scale information, Dywave employs an auxiliary reconstruction loss. A lightweight decoder reconstructs the original wavelet coefficients from the fused embeddings. The total loss is a combination of the downstream task loss and the reconstruction loss ( $L = L_{task} + \lambda_{rec} \cdot L_{rec}$ ). This regularizes the compression to preserve both high-frequency transients and low-frequency trends.

3. Key Contributions

Identification of the Tokenization Gap: The paper highlights the lack of human-intuitive semantic units in physical sensing signals and the misalignment of uniform patching with physical event dynamics.
Dywave Framework: A novel, dynamic tokenization module that converts raw signals into compact, event-aligned tokens using wavelet-based hierarchical decomposition and saliency-driven anchor selection.
Empirical Validation: Extensive evaluation across five real-world datasets (covering HAR, stress assessment, and object detection) demonstrates superior performance and efficiency.

4. Experimental Results

Dywave was evaluated on five datasets (Ego4D, MOD, PAMAP2, RWHAR, WESAD) using Transformer and Mamba2 backbones.

Accuracy: Dywave outperforms state-of-the-art methods (PatchTST, MedFormer, DropPatch, etc.) by up to 12% in accuracy on HAR tasks, particularly when using Mamba2 backbones.
Efficiency: It reduces input token lengths by up to 75% across mainstream sequence models, significantly lowering computational overhead.
Token Distribution: Unlike fixed-patch methods where token count correlates linearly with signal length, Dywave's token count is governed by semantic structure. For example, on Ego4D (30s sequences), Dywave uses significantly fewer tokens than PatchTST while achieving higher accuracy.
Robustness:
- Domain Shift: Dywave maintains higher accuracy under cross-domain generalization (e.g., seismic to audio) and head-only finetuning, where fixed-patch methods suffer severe degradation.
- Noise: Dywave exhibits greater robustness to Gaussian noise injection, with performance degrading more gradually than baselines.
- Sequence Length: It generalizes well across varying sequence lengths (5s to 30s) without re-tuning hyperparameters, whereas fixed-patch methods require extensive grid searches.
Micro-Activity Decomposition: Qualitative analysis on Ego4D shows Dywave can automatically decompose coarse activities (e.g., "cleaning") into fine-grained micro-activities (e.g., "reaching," "wiping"), aligning boundaries with perceptible motion changes.

5. Significance and Claims

The authors claim that Dywave addresses a fundamental bottleneck in human-centric sensing: the inability of current tokenization schemes to respect the physical structure of sensing data. By shifting from static, uniform windows to dynamic, event-aligned tokenization, the framework enables:

Physically Grounded Representations: Tokens correspond to distinct physical events rather than arbitrary time slices.
Adaptive Efficiency: Computational resources are allocated based on information density, reducing latency and memory usage on edge devices.
Improved Generalization: The method learns transferable representations that are less dependent on domain-specific hyperparameter tuning.

The paper concludes that adaptive, physics-informed input representations are essential for scalable and semantically meaningful learning in IoT systems. It positions Dywave not just as a performance booster, but as a principled approach to bridging the gap between continuous sensor streams and discrete deep learning inputs. The authors note that while Dywave introduces preprocessing overhead, this is amortized by substantial reductions in backbone computation, making it particularly suitable for resource-constrained, long-duration sensing scenarios.