🤖 machine learning

ProtoSSL: Interpretable Prototype Learning from Unlabeled Time-Series Data

ProtoSSL is a novel framework that learns interpretable, projection-based prototypes from unlabeled time-series data via self-supervised motif discovery, which can then be efficiently aligned to downstream tasks to achieve superior label efficiency and human-preferred explanations compared to supervised baselines.

Original authors: Steven Song, Sahil Sethi, Brett Beaulieu-Jones, Robert L. Grossman

Published 2026-05-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Steven Song, Sahil Sethi, Brett Beaulieu-Jones, Robert L. Grossman

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

Imagine you are trying to teach a computer to recognize different types of heartbeats (ECGs) or sounds (audio). Usually, to do this well, you need two things: a massive library of labeled examples (e.g., "this heartbeat is normal," "this one is dangerous") and a way to explain why the computer made its decision.

Deep learning models are great at finding patterns, but they are often "black boxes." You can't easily see what they are looking at. On the other hand, "prototype" models are like a detective who says, "I think this is a heart attack because it looks just like this specific example of a heart attack I saw before." These are easy to understand, but they usually require a huge amount of labeled data to learn those specific examples in the first place. If you change the task (e.g., from detecting heart attacks to detecting a different condition), you often have to start over and relearn everything from scratch.

Enter ProtoSSL: The "Universal Pattern Collector."

The authors of this paper created a new system called ProtoSSL that solves these problems by splitting the learning process into two distinct steps. Think of it like a librarian who first organizes books by their physical shape and color, and only later decides which books belong to which genre.

Step 1: The "Shape Collector" (Unlabeled Learning)

First, ProtoSSL looks at a massive pile of unlabeled data (thousands of heartbeats or sounds with no labels attached). It doesn't care about the diagnosis yet; it just wants to find recurring patterns.

The Analogy: Imagine a child playing with a box of mixed-up LEGO bricks. The child doesn't know what a "car" or a "house" is yet. They just sort the bricks by color, size, and shape. They build a "bank" of interesting, reusable brick patterns.
What ProtoSSL does: It uses a self-supervised method to find these recurring "motifs" (patterns) in the data and stores them in a Prototype Bank. It learns, "Oh, this specific squiggle happens often," or "This specific rhythm repeats." Crucially, it does this without knowing what these patterns mean medically or linguistically.

Step 2: The "Label Assigner" (Task Adaptation)

Once the library of patterns is built, you can now use it for a specific job. You bring in a small amount of labeled data (e.g., "these squiggles mean 'heart attack'").

The Analogy: Now, the librarian takes the pre-sorted LEGO bricks and says, "Okay, for this specific project, the red 2x4 bricks will represent 'cars,' and the blue 2x2 bricks will represent 'houses.'"
What ProtoSSL does: It efficiently matches the patterns it already found to the new labels. It asks, "Which of my pre-learned patterns best represents a 'heart attack'?" It then "grounds" these patterns by pointing to the actual real-world examples in the data that match them.

Why is this a big deal?

1. It's incredibly efficient with data.
Usually, to train a smart, explainable model, you need thousands of labeled examples. ProtoSSL can do a great job with as few as 256 labeled examples.

The Metaphor: Instead of needing a whole library of labeled books to teach a student, you just need a few examples to show them which pre-sorted chapters belong to which subject. The heavy lifting of finding the chapters was already done.

2. It's reusable.
If you train a standard model on heart data, it can't easily be used for audio data. But because ProtoSSL learns the shapes of the data first, you can take that same "Pattern Bank" and use it for completely different tasks (like switching from heartbeats to voice recognition) without retraining the whole system.

3. Humans actually like the explanations.
The researchers tested this with medical students. They showed them explanations generated by ProtoSSL and explanations from standard, label-heavy models.

The Result: Even though the standard models were sometimes slightly better at getting the answer right, the medical students preferred the ProtoSSL explanations. They felt the examples ProtoSSL picked were "better examples" of the diagnosis.
The Metaphor: A standard model might say, "This is a heart attack because of this tiny, weird spike." ProtoSSL says, "This is a heart attack because it looks exactly like this classic, textbook heart attack." Humans find the textbook example much more convincing and trustworthy.

Summary

ProtoSSL is a framework that learns what things look like before it learns what things mean. By separating the discovery of patterns from the assignment of labels, it creates a reusable, highly efficient, and human-friendly system that works well even when you have very few labeled examples. It works for heart data (ECGs) and has also been shown to work for audio data, proving it's a flexible tool for time-series analysis.

Technical Summary: ProtoSSL

Problem Statement

In time-series domains such as clinical waveform analysis and audio classification, deep neural networks often achieve high predictive performance but lack interpretability. While projection-based prototype networks (e.g., ProtoPNet) offer inherent interpretability by grounding predictions in "this looks like that" reasoning—comparing inputs to representative training examples—existing approaches suffer from two critical limitations:

Label Dependence: Prototypes are learned via supervised objectives tied to specific class structures, requiring large labeled datasets.
Lack of Reusability: Because prototypes are optimized for a specific task, they cannot be easily transferred to new downstream tasks or datasets with different label spaces, necessitating retraining from scratch.

Conversely, self-supervised learning (SSL) can learn transferable structures from unlabeled data but typically operates in latent space, failing to produce human-inspectable, projection-grounded explanations. The central challenge is how to learn prototypes that are both transferable across tasks and inherently interpretable without relying on task-specific labels during the initial learning phase.

Methodology: ProtoSSL Framework

ProtoSSL addresses this by decoupling motif discovery from label alignment. The framework operates in two distinct phases:

1. Self-Supervised Pretraining (Motif Discovery)

ProtoSSL learns a reusable bank of prototypes from unlabeled time-series data.

Architecture: An encoder $f_\theta$ maps input time series $x$ to a representation $h$ . A set of learnable prototypes $P$ is defined.
Activation: Prototype activations $a(x)$ are computed as the similarity between the encoder output and each prototype $p_k$ .
Objective: A self-supervised contrastive objective (SimCLR/NT-Xent) is applied directly to the projected prototype activations $z = r_\phi(a(x))$ . This forces the model to represent inputs as weightings of prototypes based on data consistency across augmented views, without any label supervision.
Regularization: KoLeo regularization is applied to the prototype vectors to ensure they are spread out in the latent space, preventing collapse.
Outcome: A "prototype bank" $P$ that captures recurring, transferable patterns (motifs) in the data distribution, independent of any specific downstream task.

2. Target Adaptation (Label Alignment)

Once a downstream task with labeled data becomes available, the pretrained prototypes are aligned to the new task without retraining the prototype bank.

Assignment: A novel, training-free, non-parametric assignment procedure maps prototypes from the bank to downstream labels. This is formulated as a rectangular linear assignment problem (LAP), solvable in polynomial time.
- An association score matrix $Q$ is computed based on prototype activation statistics (comparing activation on positive vs. negative samples for each label, normalized by variance).
- The LAP solver assigns the most discriminative prototypes to each label slot.
Projection: The assigned prototypes are projected onto real, representative training examples from the target dataset that possess the corresponding label. This step grounds the prototypes in concrete signal segments, preserving interpretability.
Classification: A logistic regression classifier is trained on the projected prototype activations. The encoder and prototypes can remain frozen (linear probing) or be fine-tuned with label supervision.

Key Contributions

Novel Framework: ProtoSSL is the first framework to learn projection-grounded, interpretable prototypes from unlabeled data and adapt them to downstream tasks via an explicit assignment and projection procedure.
Separation of Concerns: By separating motif discovery (self-supervised) from label alignment (assignment), the framework enables prototype reuse across datasets and label spaces without relearning the core representation.
Label Efficiency: The framework significantly improves label efficiency, remaining effective with as few as 256 labeled examples, outperforming supervised prototype baselines in low-data regimes.
Human Interpretability: A human evaluation study demonstrates that prototypes learned without task labels are rated as more representative and acceptable examples of diagnoses than those learned via direct label supervision, despite the latter having higher predictive accuracy in some cases.
Generalizability: The approach extends beyond ECG to audio classification, validating the separation of motif discovery and label alignment across distinct time-series domains.

Experimental Results

The authors evaluated ProtoSSL across six ECG datasets and three audio datasets.

ECG Performance:
- Full Data: With fine-tuning, ProtoSSL outperforms supervised prototype baselines (SupProto Direct) and achieves competitive performance against blackbox foundation models.
- Low Data: In low-data regimes (down to 256 samples), ProtoSSL (frozen probing) consistently outperforms both supervised baselines and blackbox models. Notably, ProtoSSL with only 4,000 samples matches the performance of supervised baselines trained on full datasets (64,000+ samples) on certain tasks.
- Generalization: ProtoSSL generalizes well to out-of-domain tasks where label spaces differ, whereas label-supervised pretraining (SupProto HEEDB) fails to transfer effectively to tasks with non-overlapping labels.
Audio Performance:
- ProtoSSL pre-trained on AudioSet transfers effectively to UrbanSound8K, IEMOCAP, and VoxCeleb1. It outperforms supervised baselines on tasks with minimal label overlap, demonstrating the robustness of the motif discovery phase.
Human Evaluation:
- In a blinded study with medical students, ProtoSSL prototypes were rated significantly higher in "acceptability" (91.4% vs. 66.4%) and explanation quality (82.9% vs. 67.9%) compared to directly supervised prototypes. This suggests that self-supervised discovery yields patterns that are more consistently perceived as representative clinical examples.
Efficiency:
- The assignment step is computationally efficient, solving the linear assignment problem in seconds, whereas alternative parametric assignment methods (e.g., adapted ProtoPool) are an order of magnitude slower.

Significance and Claims

The paper claims that ProtoSSL fundamentally shifts the paradigm for interpretable time-series modeling by demonstrating that interpretable prototypes can be learned before the downstream label space is known.

Decoupling: It proves that the "this looks like that" reasoning mechanism does not require end-to-end label supervision to discover useful patterns.
Reusability: It establishes a pathway for creating reusable, interpretable model components that can be rapidly adapted to new clinical or industrial tasks with minimal labeled data.
Quality of Evidence: The authors posit that self-supervised motif discovery may yield more representative case-based evidence than supervised learning, which might prioritize discriminative features that are less intuitively "prototypical" to human experts.

The work positions ProtoSSL as a foundation for studies in data-scarce domains (such as medicine) where obtaining large labeled datasets is difficult, yet interpretability is essential for clinical trust and adoption.