TLDR: This research introduces “Learning Without Augmenting” (LWA), a novel self-supervised learning method for time series data that bypasses the need for traditional data augmentations. Instead, LWA generates diverse data views by projecting time series onto orthonormal bases (like Fourier transforms) and overcomplete frames (like Gabor wavelets). By leveraging the distinct geometric biases of these projections and using latent space mappers, the method achieves significant performance improvements (up to 15-20%) on various temporal sequence tasks without increasing data diversity through handcrafted augmentations or requiring larger models.
Self-supervised learning (SSL) has become a powerful way to train models without needing a lot of labeled data. However, most current SSL methods heavily rely on what are called ‘data augmentations.’ These are essentially handcrafted transformations applied to the data to create diverse versions, which then help the model learn robust representations. Think of it like showing a model many slightly altered pictures of a cat (rotated, zoomed, color-shifted) so it learns what a cat is, regardless of these minor changes.
The challenge with these augmentations is that they often require deep, domain-specific knowledge to design effectively. If not done carefully, they can impose specific biases on the model, limiting its ability to generalize to new, unseen data. For complex data types like time series (e.g., heart rate, activity data, brain signals), designing effective augmentations is particularly difficult and can even lead to ‘model collapse,’ where the model fails to learn anything useful.
A new research paper titled “Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections” by Berken Utku Demirel and Christian Holz from ETH Zürich introduces a novel approach that sidesteps this reliance on handcrafted augmentations. Their method, which we’ll call LWA for short, generates diverse views of time series data not by altering the data itself, but by projecting it onto different mathematical spaces: orthonormal bases and overcomplete frames.
At its core, LWA uses two well-established mathematical transformations: the Fourier transform and the Gabor wavelet transform. The Fourier transform provides a global overview of a signal’s frequency content, revealing its underlying periodic patterns. The Gabor wavelet transform, on the other hand, offers a localized frequency analysis, allowing the model to zoom in on specific time intervals and capture transient events. These two transformations are complementary, meaning they capture different but equally important characteristics of the signal.
The researchers found that the representations learned from these different mathematical spaces (time domain, Fourier domain, and Gabor wavelet domain) naturally reside on distinct ‘manifolds’ – essentially, different geometric structures in the model’s internal representation space. Each manifold is shaped by the inherent biases introduced by representing the data in its respective space. By jointly leveraging the complementary geometry of these distinct manifolds, LWA achieves superior performance without needing to artificially increase data diversity through strong augmentations.
The method employs an instance discrimination task, where the model learns to distinguish between different samples across these transformed spaces. This is done using a loss function that encourages embeddings of the same sample (but from different domains) to be similar, while pushing embeddings of different samples apart. A crucial aspect of LWA is its use of lightweight ‘latent space mappers.’ These are small neural networks that learn to transform representations from the original time domain into the Fourier and Gabor wavelet domains. This allows the model to leverage the rich geometry of these alternative spaces during inference, even though the original data is only processed by a single main encoder.
The effectiveness of LWA was demonstrated across nine datasets spanning five temporal sequence tasks, including heart rate estimation, activity recognition, cardiovascular disease classification, step counting, and sleep stage classification. The results showed impressive performance gains of up to 15–20% over existing self-supervised approaches. Importantly, LWA uses fixed transformations across all datasets, unlike other methods that often require task-specific augmentations to be carefully tuned.
A key takeaway from this research is the idea that strong, hand-crafted data augmentations might not be as essential for self-supervised learning as previously thought, especially for temporal signals. The paper suggests that instead of distorting data, leveraging intrinsic geometric biases through principled mathematical transformations can be a more effective and generalizable strategy. While the work primarily focused on classification tasks, the integration of both global (Fourier) and local (Gabor) representations holds promise for forecasting tasks as well.
Also Read:
- Unlocking Intuitive Audio Manipulation with Linear Latent Spaces
- VESSA: Adapting Vision Models with Object-Centric Videos
This work opens new avenues for self-supervised learning, emphasizing mathematical structure over empirical trial and error, and paving the way for more generalizable, augmentation-free methods across various domains. You can read the full research paper here: Learning Without Augmenting: Unsupervised Time Series Representation Learning via Frame Projections.
