MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition

arXiv:2610.10245 · cs.LG, cs.HC · Submitted 2026-10-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition".

Jane: The gist The Morphological Contrastive Learning (MorphCL) framework introduces a self-supervised pretraining method that uses structure-aware grouping to inject explicit modeling of global structure into inertial-based self-supervised learning approaches,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we’re wrapping up our look at MorphCL which is Morphological Contrastive Learning for Inertial-based Human Activity Recognition.

Jane: Right, so this paper essentially introduces a way to use structure-aware grouping to force explicit modeling of global structure into those existing self-supervised learning methods we discussed earlier.

Tom: Exactly. It shifts the whole focus from just looking at individual data points to understanding the larger motion patterns that exist across time and different activities.

Jane: And they show that by doing this, you can see a real jump in performance, up to fifteen percentage points on F1-scores when you combine it with other self-supervised learning frameworks.

Tom: That’s a big number. And they even show their pretraining encoders matching foundation models trained on four thousand six hundred times less data for things like the WISDM and CAPTURE-twenty-four datasets <ref:2610.10245#pg3>.

Jane: It suggests that this way of organizing the data, this morphological structure, is a really powerful inductive bias that helps the AI learn much better when it’s starting with limited training examples.

Tom: And what I find interesting is how they define their loss function there—that group-based contrastive objective. It specifically makes sure the segments agreeing are truly sharing motion patterns rather than just being similar in a local sense.

Jane: So, for someone listening right now, it means that if you’re working with motion sensors and want better results without needing massive amounts of labeled data immediately, you should look at how structure can be explicitly modeled.

Tom: Exactly. It shows that focusing on the global patterns of movement can actually complement the local similarity objectives in contrastive learning much more effectively.

Jane: But we have to keep an eye on how stable this works when you try to deploy it—the authors did show it’s pretty robust across different sampling techniques, which is something Meng will be interested in later.

Tom: Right, so we’ve seen the mechanics and the numbers behind MorphCL. Next up, we're going to dig into what these results actually mean for deploying this technology in real-world scenarios.

Conclusion: Tom: So, we're wrapping up our look at MorphCL which is Morphological Contrastive Learning for Inertial-based Human Activity Recognition.

Tom: Exactly. It shifts the whole focus from just looking at individual data points to understanding the larger motion patterns that exist across time and different activities.

Jane: And they show that by doing this, you can see a real jump in performance, up to fifteen percentage points on F1-scores when you combine it with other self-supervised learning frameworks.

Tom: That improvement is up to fifteen percentage points on F1-scores when they combine MorphCL with other SSL frameworks.

Jane: And the paper shows their pretraining encoders actually match representations from a foundation model that was trained on four thousand six hundred times more data for datasets like WISDM and CAPTURE-twenty-four.

Tom: That suggests that this way of organizing the data, this morphological structure, is a really powerful inductive bias that helps the AI learn much better even when you start with limited training examples.

Jane: And what they do to get there is introduce three stages: motif discovery, clustering of domain-specific features, and then a group-based contrastive objective.

Tom: That's the mechanism we need to look at for a second part of our discussion. The first stage is motif discovery, which uses multi-dimensional matrix profiling to find prominent repeated patterns or motion primitives within an inertial time series.

Jane: So they are essentially searching for these recurring motion patterns within the raw sensor data itself, identifying up to m distinct motif patterns per chunk and extracting r instances of each that stay within a set distance threshold sigma.

Tom: And then they move to morphological clustering where these extracted motif windows are represented as feature vectors, z i equals Phi(c i), and those vectors are clustered using HDBSCAN based on density.

Jane: So this creates a mapping to K, the number of morphological clusters discovered, resulting in a filtered dataset D that is labeled based on these detected motifs.

Tom: The final step is the Morphological Contrastive Learning objective, which they define as Lmorph equals one over B summed over i minus log P p in P(i) times exp of sigma summed over j not in i exp of similarity tau between u i and u j.

Jane: That loss function is group-based, so it encourages segments that share similar motion patterns to agree on each other rather than just focusing on isolated pairs.

Tom: So for someone listening right now, it means that if you’re working with motion sensors and want better results without needing massive amounts of labeled data immediately, you should look at how structure can be explicitly modeled before you even start your training.

Jane: It shows that focusing on the global patterns of movement can actually complement the local similarity objectives in contrastive learning much more effectively than just using standard augmentation techniques.

Tom: But we have to consider what this means for deployment, because the authors did show it’s pretty robust across different sampling techniques, which is something Meng will be interested in later.

Jane: Right, so we've seen the high-level idea and the performance claims behind MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition. Now let's go into what this actually means for future work and real-world application.

Lalam: The authors are really emphasizing that this isn't just a tweak to existing contrastive learning. They're building an explicit structural layer on top of the data representation itself.

Lu: I agree. What's interesting is how they treat the motion primitives as distinct entities before clustering them into morphology. That separation between finding those small patterns and then grouping them is a solid way to inject that global knowledge.

Meng: From my side, if we can use this pretraining to get better initial weights, it means we spend less time collecting massive amounts of labeled activity data just to get the model off the ground.

Tom: Exactly. And those four thousand six hundred times less data comparison they showed? That’s the part that gets me. It suggests this structure is genuinely learning useful features from raw sensor noise, not just memorizing labels.

Jane: So, for someone who only listens to this show, it means we're moving toward AI systems that inherently understand the *shape* of human movement before they try to classify what activity is happening.

Lu: If this method scales well with different types of motion sensors—not just IMUs but maybe even other time-series data—that would be a big expansion for representation learning.

Tom: Right, so we’ve seen the mechanics and the numbers behind MorphCL. Next up, we're going to dig into what these results actually mean for deploying this technology in real-world scenarios.

Marius Bock, Yuwei Zhang, Juergen Gall, Michael Moeller, Kristof Van Laerhoven, Cecilia Mascolo

University of Bonn · University of Cambridge · Lamarr Institute for Machine Learning and Artificial Intelligence

cs.LG, cs.HC

Submitted: 2026-10-07

Updated: 2026-10-07

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: The gist The Morphological Contrastive Learning (MorphCL) framework introduces a self-supervised pretraining method that uses structure-aware grouping to inject explicit modeling of global structure

Key concepts

Motif Discovery: This stage uses multi-dimensional matrix profiling on inertial time series to find prominent, repeated movement patterns called 'motifs.' It identifies up to 'm' distinct motifs per chunk and extracts multiple instances of each that are close enough in distance. A greedy masking strategy is used to ensure the discovered motifs cover diverse motion types.
Morphological Clustering: The extracted motif windows are converted into feature vectors using statistical features common in activity recognition. These vectors are then clustered using HDBSCAN, a density-based algorithm, to group similar motion patterns into 'morphological clusters.' This creates a labeled dataset where each motif instance belongs to a specific morphological category.
Morphological Contrastive Learning (MORPHCL): This is the core self-supervised loss function. Instead of comparing isolated pairs, it uses a group-based variant of Supervised Contrastive Loss (SCL). This loss encourages the model to pull together segments that share similar motion patterns within the same morphological cluster, forcing the embedding space to reflect global structural consistency.
Global Structure Modeling: The central idea is that human movement data contains recurring patterns across time and activities. Previous methods often ignore this structure. MorphCL explicitly models this by identifying and clustering these underlying motion primitives, allowing the model to learn representations based on the overall shape of the movement rather than just individual data points.

Terminology

Contrastive Learning — 1× in this paper · explained in 7 episodes

Summary

The gist The Morphological Contrastive Learning (MorphCL) framework introduces a self-supervised pretraining method that uses structure-aware grouping to inject explicit modeling of global structure into inertial-based self-supervised learning approaches, substantially improving downstream results by up to 15 percentage points in F1-score

Introduction and Motivation

Activities and movements performed during daily life encode rich information about an individual’s behavioral patterns, physiological state, and habitual routines The automatic identification of such behaviors, commonly referred to as Human Activity Recognition (HAR), has therefore attracted growing interest across a range of health-related research fields Inertial Measurement Units (IMUs), now ubiquitous in consumer wearables and mobile devices, have emerged as a particularly practical sensing modality for this purpose, yielding discriminative representations well-suited to activity recognition systems While clever pretext tasks such as contrasting differently augmented views of input signals have shown to perform well on various downstream tasks, such methods typically rely on instance-level discrimination, treating each segment as independent This assumption is often violated in IMU data, where recurring motion primitives are shared across activities and subjects, making instance-level objectives poorly aligned with the underlying structure The authors argue that a key missing component in current SSL approaches for HAR is the explicit modeling of global structure in motion data because human movement exhibits recurring patterns that appear across time, activities, and individuals

MorphCL Methodology

MorphCL is a self-supervised pretraining framework that can serve as a plug-in solution to inject explicit modeling of global structure into existing SSL approaches The framework consists of three stages: (i) motif discovery, (ii) clustering of domain-specific features, and (iii) a group-based contrastive objective

Motif Discovery: This stage involves using multi-dimensional matrix profiling to identify prominent repeated patterns or motion primitives within an inertial time series They perform motif discovery independently on temporally contiguous chunks, identifying up to m distinct motif patterns per chunk and extracting r instances of each which do not exceed a pre-defined motif distance threshold σ A greedy masking strategy is adopted to promote diversity among the discovered motifs
Morphological Clustering: Extracted motif windows are represented as feature vectors zi = Φ(ci) ∈ R F where Φ(·) concatenates statistical features commonly used in classical machine-learning-based activity recognition These feature vectors are then clustered using HDBSCAN, a density-based clustering algorithm, to induce a mapping κ: R F → K ∪ ∅ where K denotes the number of morphological clusters discovered in Z This results in a motif-filtered and morphologically labeled dataset D = (wi, yi)
Morphological Contrastive Learning (MORPHCL): The loss is defined as Lmorph = 1/B Σ i∈B - log P p∈P(i) exp (Σ j∈B i exp (sim(ui,uj) τ) This group-based variant of the Supervised Contrastive Loss (SCL) encourages agreement between segments sharing similar motion patterns rather than isolated pairs

Experimental Results and Comparisons

The authors demonstrate that including explicit modeling of morphological structure can improve downstream results if combined with existing SSL frameworks by up to 15% in F1-score They show that MorphCL-pretrained encoders match or surpass foundation models trained on 4600× less data while trained on the WISDM and CAPTURE-24 datasets

**: In comparison with existing contrastive learning methods like REBAR and RelCon, MorphCL benefits most substantially from the addition of MorphCL, suggesting that while improved pair retrieval is valuable, it does not fully compensate for the lack of global structural awareness in the embedding space The combined objective ((c) MorphCL + MTL) yields both structure and separability, with clusters being compact and well-delineated Furthermore, qualitative analysis reveals that the global structure becomes more coherent, with samples organized into smoother manifolds that reflect morphological consistency The results are robust to different sampling techniques, showing stability even when not using oversampling and balanced batches Additionally, the group-based loss formulation results in the on average highest F1-scores across the different downstream experiments The framework is stable across different motif discovery hyperparameters, with the chosen configuration of σ = 1.0 and r = 100 returning most stable results for motifs of length 0.

Improvements for AI systems

textbf Morphological Contrastive Learning for Global Structure Modeling in Inertial Data Development of a Group-Based Contrastive Objective: This improves self-supervised learning by inject[ing] explicit modeling of global structure into inertial-based SSL approaches through the construction of a group-based contrastive objective that encourages agreement between segments sharing similar motion patterns, moving beyond instance-level comparisons.
textbf Motif Discovery and Feature Clustering for Structure-Aware Pseudo-Label Generation: This component uses motif discovery to identify recurring motion primitives and then clusters these motifs using HDBSCAN on domain-specific features, resulting in a morphological clustering of the pretraining corpus that defines which windows are treated as positives (same group) and negatives (different groups).
textbf Enhanced Downstream Performance via MorphCL Loss Injection: Incorporating the MorphCL loss term into existing SSL frameworks can improve representation quality, with linear probing improvements ranging up to as much as 15% in F1-score, suggesting that domain-informed inductive bias can complement large scale training.
textbf Foundation Model Scaling Efficiency: The framework allows encoders to achieve performance competitive with a foundation model that was trained on 4600× more data, indicating that the method enables better generalization by using the available data more efficiently rather than relying solely on massive pretraining scales.

Abstract

Despite the ubiquity of sensors in wearable and mobile devices and the abundance of human movement data they generate, translating unlabeled recordings into foundational motion models remains an open challenge. Self-supervised learning (SSL) has alleviated the need for costly annotations, yet existing approaches leave the global structure of large-scale motion data largely untapped, relying on randomly sampled batches and local comparisons that become particularly problematic for in-the-wild inertial data dominated by stationary, low-variance behaviors. Here we introduce Morphological Contrastive Learning (MorphCL), a self-supervised pretraining framework that uses structure-aware grouping to inject explicit modeling of global structure into inertial-based SSL approaches. Building on two well-established pillars of motion analysis, the discovery of motion primitives, or motifs, and domain-specific feature descriptors, we show that MorphCL substantially improves linear probing and finetuning results of learned encoders by up to 15 percentage points in F1-score. In a comparison with existing foundation models, we demonstrate that MorphCL-pretrained encoders match or surpass them models in linear probing performance while trained on 4600 times less data. Qualitative analysis of the resulting embedding spaces further reveals morphologically meaningful cluster structure, with improved separation of kinematically similar activity classes.