Dataset Distillation · Preprint 2026

Closed-Form Linear-Probe Dataset Distillation for Pre-trained Vision Models CLP-DD

1Hokkaido University  ·  2University of Nevada, Reno *Equal contribution  ·  Correspondence to: Guang Li <guang@lmd.ist.hokudai.ac.jp>
Accuracy versus efficiency on ImageNet-100 with DINOv2 at IPC=1
Figure 1. Accuracy versus efficiency on ImageNet-100 with DINOv2 at IPC = 1. CLP-DD attains accuracy close to LGM with DSA at the runtime and memory of LGM without DSA. Bubble area denotes peak GPU memory.

Abstract

Dataset distillation compresses a large training set into a small synthetic set that preserves downstream training utility. While most existing methods target training networks from scratch, modern visual transfer learning often uses frozen pre-trained encoders followed by lightweight linear probing. Existing distillation methods for this setting either unroll iterative linear-probe updates with trajectory-based gradient matching, or rely on closed-form formulations originally designed for from-scratch training with neural-tangent-kernel (NTK) approximations. Neither route exploits the fact that frozen-feature linear probing admits a closed-form solution determined directly by the pre-trained features themselves, with no infinite-width approximation and no inner-loop trajectory. We propose Closed-Form Linear-Probe Dataset Distillation (CLP-DD), a bilevel formulation that computes the linear probe induced by the synthetic set with a sample-space kernel ridge solver. The synthetic images are then updated by evaluating this induced classifier on real features through a temperature-scaled softmax cross-entropy, where the classifier columns act as learned class anchors in feature space. We further show that the choice of outer objective is decisive: pairing the closed-form inner solver with a standard MSE outer loss substantially underperforms trajectory-based methods, while the discriminative outer loss closes most of the gap. On ImageNet-100 with four pre-trained backbones, CLP-DD substantially improves over LGM without DSA and approaches LGM with DSA at a fraction of the computational cost. On ImageNet-1K, CLP-DD matches or surpasses LGM with DSA on three of four backbones while running roughly 14× faster and using less than one-eighth of the GPU memory.

At a Glance

TL;DR — For frozen encoders, the linear probe has an exact closed form. CLP-DD solves it analytically in sample space and back-propagates through it with a class-anchor cross-entropy loss — no trajectory unrolling, no NTK approximation.
14×
faster than trajectory-based LGM w/ DSA
<1/8
of the peak GPU memory
68.3%
ImageNet-1K average at IPC = 1
(vs. 67.8% for LGM w/ DSA)
+2.9
points from the class-anchor outer loss over MSE

Method

Setting. A frozen pre-trained encoder $\varphi$ maps images to features; only a linear probe is trained downstream. The distillation goal is a tiny synthetic set $\mathcal{S}_{\text{syn}}$ whose induced probe generalizes on real data.

  1. Synthetic features through the frozen encoder

    The learnable synthetic images are passed through the frozen backbone to obtain a feature matrix.

    $$\mathbf{X}_{\text{syn}} \;=\; \varphi(\mathcal{S}_{\text{syn}}) \;\in\; \mathbb{R}^{N \times d}$$
  2. Inner problem: ridge-regularized linear probe

    Linear probing on frozen features is a convex quadratic problem — it does not need to be unrolled.

    $$\mathcal{L}_{\text{in}}(\mathbf{W}) \;=\; \tfrac{1}{2}\big\lVert \mathbf{X}_{\text{syn}}\mathbf{W} - \mathbf{Y}_{\text{syn}} \big\rVert_F^2 \;+\; \tfrac{\lambda}{2}\lVert \mathbf{W} \rVert_F^2$$
  3. Exact closed-form solution

    The minimizer is available analytically — no inner loop, and no infinite-width (NTK) approximation.

    $$\mathbf{W}^{\star} \;=\; \big(\mathbf{X}_{\text{syn}}^{\top}\mathbf{X}_{\text{syn}} + \lambda \mathbf{I}_d\big)^{-1}\mathbf{X}_{\text{syn}}^{\top}\mathbf{Y}_{\text{syn}}$$
  4. Sample-space kernel reformulation

    In the low-IPC regime the number of synthetic samples $N$ is far below the feature dimension $d$. The push-through identity turns a $d \times d$ solve into a much cheaper $N \times N$ solve — this is where most of the speed and memory savings come from.

    $$\mathbf{W}^{\star} \;=\; \mathbf{X}_{\text{syn}}^{\top}\big(\underbrace{\mathbf{X}_{\text{syn}}\mathbf{X}_{\text{syn}}^{\top}}_{\mathbf{K}_{\text{syn}}} + \lambda \mathbf{I}_N\big)^{-1}\mathbf{Y}_{\text{syn}}, \qquad N \ll d$$
  5. Outer objective: classifier columns as class anchors

    Instead of an MSE fit to one-hot targets, the induced probe is evaluated discriminatively on real features with a temperature-scaled softmax cross-entropy. Each column $\mathbf{w}^{\star}_j$ behaves as a learned class anchor in feature space.

    $$\mathcal{L}_{\text{meta}} \;=\; -\frac{1}{M}\sum_{i=1}^{M} \log \frac{\exp\!\big(\mathbf{x}_{\text{real},i}^{\top}\mathbf{w}^{\star}_{y_i}/\tau\big)}{\sum_{j}\exp\!\big(\mathbf{x}_{\text{real},i}^{\top}\mathbf{w}^{\star}_{j}/\tau\big)}$$

Why it is cheap. The inner problem is solved once, exactly, as an $N \times N$ linear system — no unrolled optimizer states to store and no backward pass through a trajectory. This is what buys the 14× speedup and the 8× memory reduction reported below.

Efficiency

Per-iteration cost on ImageNet-100 with DINOv2 at IPC = 1. CLP-DD runs at the cost of DSA-free LGM while reaching higher accuracy than LGM with DSA on this backbone setting.

Runtime (seconds, lower is better)

LGM w/ DSA57.49 s
LGM (no DSA)4.25 s
CLP-DD (ours)4.07 s
≈ 14× faster than LGM w/ DSA

Peak GPU memory (MB, lower is better)

LGM w/ DSA19,320 MB
LGM (no DSA)2,334 MB
CLP-DD (ours)2,361 MB
< 1/8 of the memory of LGM w/ DSA

Results

Table 1a. ImageNet-100, IPC = 1. Linear-probe accuracy (%) across four frozen backbones.
MethodDINOv2MoCo-v3CLIPEVA-02Average
Random74.8±2.661.3±2.657.6±1.564.4±2.564.5±2.5
Centroid86.9±0.377.7±0.177.1±0.180.9±0.280.6±0.2
Neighbor86.0±0.277.1±0.267.8±0.278.8±0.177.4±0.2
LGM (no DSA)82.6±0.259.5±0.558.4±0.373.9±0.268.6±0.4
LGM w/ DSA91.4±0.183.3±0.184.6±0.188.6±0.187.0±0.1
CLP-DD (ours)89.2±0.183.9±0.179.2±0.187.6±0.085.0±0.1
Full dataset95.1±0.189.3±0.292.4±0.094.0±0.192.7±0.1
Table 1b. ImageNet-1K, IPC = 1. CLP-DD matches or surpasses LGM w/ DSA on three of four backbones — at a fraction of the cost.
MethodDINOv2MoCo-v3CLIPEVA-02Average
Random50.3±0.538.7±0.531.7±0.437.7±0.639.6±0.5
Neighbor67.7±0.156.3±0.138.7±0.150.0±0.153.2±0.1
Centroid69.5±0.057.4±0.153.8±0.158.0±0.059.6±0.1
LGM w/ DSA75.0±0.163.2±0.062.9±0.070.3±0.167.8±0.1
CLP-DD (ours)75.6±0.163.4±0.162.4±0.171.2±0.168.3±0.1
Full dataset83.0±0.176.5±0.078.7±0.081.7±0.080.0±0.1
Table 2. WaterBirds, IPC = 1. Robustness under background shift. CLP-DD is the only efficient method that stays close to LGM w/ DSA.
MethodDINOv2MoCo-v3CLIPEVA-02Average
Random57.1±8.667.4±5.771.9±4.559.8±13.364.0±8.0
Neighbor67.3±5.845.2±11.474.6±6.574.3±3.365.3±6.8
Centroid65.7±3.462.0±5.269.8±8.664.6±7.965.5±6.3
LGM58.3±3.263.4±1.664.1±2.264.8±1.862.6±2.2
LGM w/ DSA82.1±2.977.8±0.077.9±0.278.0±0.879.0±1.0
CLP-DD (ours)80.0±2.378.5±0.475.4±0.379.6±0.478.3±0.8
Full dataset95.5±0.174.3±0.186.0±0.190.4±0.286.5±0.1

The outer objective is decisive

Table 5. Ablation on the outer loss (ImageNet-100, IPC = 1). Replacing MSE on one-hot targets with the temperature-scaled class-anchor cross-entropy adds +2.9 points on average — and +5.1 on CLIP.
Outer objectiveDINOv2MoCo-v3CLIPEVA-02Average
MSE (w/o class anchor)88.3±0.082.1±0.174.1±0.184.1±0.182.1±0.1
Class-anchor CE (ours)89.2±0.183.9±0.179.2±0.187.6±0.085.0±0.1

BibTeX

Please cite our work if you find it useful.
@article{peng2026clpdd,
  title   = {Closed-Form Linear-Probe Dataset Distillation for Pre-trained Vision Models},
  author  = {Peng, Bincheng and Li, Guang and Liu, Ping and Ogawa, Takahiro and Haseyama, Miki},
  journal = {arXiv preprint arXiv:2605.07194},
  year    = {2026}
}