When Transformers Sing: Adapting SpectralKD for Text-Based Knowledge Distillation

Whereas engaged on my Data Distillation downside for intent classification, I confronted a puzzling roadblock. My setup concerned a trainer mannequin, which is RoBERTa-large (finetuned on my intent classification), and a scholar mannequin, which I used to be attempting to coach with out shedding an excessive amount of accuracy in comparison with the trainer.

I experimented with a number of mapping strategies, connecting each 2nd layer to the coed layer, averaging two trainer layers into one, and even assigning customized weights like giving (0.3 to l1 and 0.7 to l2). However it doesn’t matter what mixture I attempted, the trainer’s accuracy by no means matched the coed mannequin.

That’s once I began exploring the way to map probably the most informative layers to my scholar mannequin in order that the coed can maximize its efficiency. I needed a option to quantify which layer of the trainer mannequin actually issues for distillation.

In that search, I stumbled upon an enchanting paper—”SpectralKD: A Unified Framework for Interpreting and Distilling Vision Transformers via Spectral Analysis,” which tackled the same downside however within the picture area. The authors used a spectral evaluation strategy (Spectral KD) to extra intelligently align the trainer and scholar fashions.

Curious, I made a decision to adapt the thought to textual content knowledge – and BOOM!!!, it truly labored! For the primary time, my scholar mannequin began pondering virtually like its trainer.

Right here’s the layer depth graph of my fine-tuned RoBERTa-large mannequin. Based mostly on the spectral insights, I chosen layers 1–9 and 21–23 for my scholar mannequin throughout information distillation, those carrying the richest info.

I can’t share my dataset or code for confidentiality causes, however I’ll stroll you thru how the paper’s image-based strategy impressed my text-based adaptation, and how one can take into consideration doing the identical.

---

Behind the Scenes: How FFT Reveals a Mannequin’s Spectral Soul

So, let’s begin with spectral depth, and slowly dive into the actual magician right here: the Quick Fourier Rework (FFT).

Within the spectralKD paper, the authors introduce a framework that helps us to see Imaginative and prescient Transformer(ViTs), not simply what they’re predicting, but in addition how the knowledge flows within the layers. As a substitute of counting on instinct or visualisation, they use spectral evaluation, a approach to measure the frequency richness of the mannequin’s inner representations.

Think about every Transformer layer because the musician in an orchestra, some layers play excessive notes(superb particulars), whereas others play low notes(broad options). The FFT helps us to hear to every participant’s music individually and filter out which one is having the strongest melodies, i.e., probably the most information-rich indicators.

Step 1: Characteristic maps, The uncooked materials

B is batch measurement
C is variety of channels and,
H,W is the spatial peak and width.

Step 2: Making use of the fourier Rework

The authors apply a 1-dimensional FFT alongside the channel dimension to translate these real-valued activations into the frequency area:
F(X)=FFT(X)

This implies:
For each spatial location (b, h, w), a 1D FFT is computed throughout all channels.
The result’s a complex-valued tensor (since FFT outputs actual + imaginary components).
F(X) subsequently tells us how a lot of every frequency is current in that layer’s illustration.

And in case you’re questioning, “Why FFT although?” — maintain that thought.
As a result of later on this weblog, we’re going to uncover precisely why FFT is the proper instrument to measure a mannequin’s internal depth.

Step 3: measuring frequency power

Re(F(X)) is the actual half,
Im(F(X)) is the imaginary half.

Step 4: Averaging throughout the map

Now we wish to summarize this depth throughout all positions within the layer:

This step tells us the common depth of the only channel

After which you may merely do common of every channels. Voilà! Now you will have the spectral depth of the only layer of the Imaginative and prescient Transformer.

Peeking into the Frequency Realm: The Fourier Lens of SpectralKD

Let’s look into the Quick Fourier Rework:

Xₖ is the enter sequence (your sign, function, or activation sample).
xₙ is the frequency part on the frequency index.
N is the variety of factors within the sequence (i.e., variety of channels or options).

Every time period e⁻ʲ²πᵏⁿ/ᴺ acts as a rotating phasor, a tiny complicated wave spinning by way of the sign house, and collectively, they kind some of the lovely concepts in sign processing.

.OMG! What simply occurred right here? Let me break it down.

Whenever you multiply your hidden activations xₙ (say, throughout channels or function dimensions) by this phasor, you’re primarily asking:

“Hey, layer, how a lot of the k-th kind of variation do you include in your representations?”

Every frequency okay corresponds to a definite sample scale throughout the function dimensions.

Decrease okay values seize broad, easy semantic buildings (like topic-level context), whereas increased okay values seize speedy, fine-grained variations (like token-level nuances or syntactic indicators).

Now right here’s the enjoyable half: if some layer resonates with a selected frequency sample, the multiplication of the Fourier Rework aligns completely, and the sum within the Fourier system produces a robust response for that okay.

If not, the rotations cancel out, which means that frequency doesn’t play an enormous position in that layer’s illustration.

So, the Fourier Rework isn’t including something new; it’s simply discovering out how our layer encodes info throughout totally different scales of abstraction.

It’s like zooming out and realizing:

• Some layers hum quietly with easy, conceptual meanings (low frequencies),
• Others buzz with sharp, detailed interactions between tokens (excessive frequencies).

The FFT principally turns a layer’s hidden states right into a frequency fingerprint — a map of what sorts of knowledge that layer is specializing in.

And that’s precisely what SpectralKD makes use of to determine which layers are truly doing the heavy lifting throughout information distillation.

For those who nonetheless want the visualization and extra instinct of the Fourier rework, you may simply undergo the 3Blue1Brown Video, “But what is the Fourier Transform? A visual introduction.”

From Imaginative and prescient to Language: How Spectral Depth Guided My Intent Classifier

Let a layer activation tensor be:

the place:

• N = variety of samples (batch measurement)
• L = sequence size (variety of tokens/time steps)
• H = hidden dimension (variety of channels/options produced by the layer)

Every Pattern i has an activation matrix Xᵢ ∈ Rᴸ ˣ ᴴ (sequence positions x hidden options)

Now once more, you may compute the FFT of that Xᵢ after which measure the frequency size utilizing the actual and imaginary parts and common out throughout the channels, after which for every layer.

Frequency size:

Frequency throughout channels:

Frequency throughout a layer:

Right here, Okay is the variety of bins retained.

Conclusion

Their evaluation exhibits two main insights:

1. Not all layers contribute equally. In uniform transformer architectures, just a few early and last layers present robust spectral exercise, the true “hotspots” of knowledge move.
2. Totally different transformer sorts, related melodies. Regardless of architectural variations, each hierarchical and uniform transformers share surprisingly related spectral patterns, hinting at a common approach these fashions study and characterize information.

Constructing on these findings, SpectralKD introduces a easy, parameter-free information distillation (KD) technique. By selectively aligning the spectral conduct of early and last layers between a trainer and a scholar mannequin, the coed learns to mimic the trainer’s spectral signature, even in intermediate layers that have been by no means explicitly aligned.

The outcomes are hanging within the paper: the distilled scholar (DeiT-Tiny) doesn’t simply match efficiency on benchmarks like ImageNet-1K, it additionally learns to assume spectrally just like the trainer, capturing each native and international info with outstanding allegiance.

Finally, SpectralKD bridges interpretability and distillation, providing a contemporary option to visualize what occurs inside transformers throughout studying. It opens a brand new line of analysis, the authors name “distillation dynamics”, a journey into how information itself flows, oscillates, and harmonizes between trainer and scholar networks.

References

Core Spectral & Transformer Foundations

• Vaswani, A. Attention Is All You Need. NeurIPS, 2017.
• Dosovitskiy, A. An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale. arXiv preprint arXiv:2010.11929, 2020.
• Raghu, M., Unterthiner, T., Kornblith, S., Zhang, C., & Dosovitskiy, A. Do Vision Transformers See Like Convolutional Neural Networks? NeurIPS, 2021.
• Han, K. et al. A Survey on Vision Transformer. IEEE TPAMI, 2022.

Interpretability & Spectral Evaluation

• Chefer, H., Gur, S., & Wolf, L. Transformer Interpretability Beyond Attention Visualization. CVPR, 2021.
• Yeh, C. et al. AttentionViz: A Global View of Transformer Attention. IEEE TVCG, 2023.
• Zeng, J. et al. Peeling Back the Layers: Interpreting the Storytelling of ViT. ACM Multimedia, 2024.

Data Distillation & Mannequin Compression

• Hinton, G. Distilling the Knowledge in a Neural Network. arXiv preprint arXiv:1503.02531, 2015.
• Phuong, M., & Lampert, C. Towards Understanding Knowledge Distillation. ICML, 2019.
• Park, W. et al. Relational Knowledge Distillation. CVPR, 2019.
• Chandrasegaran, K. et al. Revisiting Label Smoothing and Knowledge Distillation Compatibility: What Was Missing? ICML, 2022.
• Huang, T. et al. Knowledge Distillation from a Stronger Teacher. NeurIPS, 2022.
• Pham, C. et al. Frequency Attention for Knowledge Distillation. WACV, 2024.
• Fan, J. et al. ScaleKD: Strong Vision Transformers Could Be Excellent Teachers. arXiv preprint arXiv:2411.06786, 2024.
• Son, S. et al. The Role of Masking for Efficient Supervised Knowledge Distillation of Vision Transformers. ECCV, 2025.

SpectralKD Core Paper