Why it’s here
The record holds one gradient-free training line — evolution strategies, under SOTA-154 — and nothing at all on local-learning methods: no forward-forward, no target propagation, no forward gradients. That is a whole family of answers to “what if not backpropagation”, and this is its highest-traffic recent entry point.
It is also a clean instance of machinery moving between subfields. The objective is a variational diffusion ELBO; the thing being diffused is a class label embedding; and the network is a classifier.
What it does
Take a classification pair (x, y). Embed y into R^m with a trainable
matrix and run a variance-preserving Ornstein-Uhlenbeck process on it, so
z_0 … z_T is a noisy trajectory from the label to Gaussian noise. Train each
block û_t independently to predict the clean label embedding from
(z_{t-1}, x) under an L2 loss, with a cross-entropy term on the final
readout.
At inference the blocks run in sequence: each takes the noisy embedding the previous one produced and steps it toward its own prediction. So the forward pass looks like a residual network with noise — but it was never trained as one. At training time there is no forward propagation and no backward propagation across blocks, which is where the name comes from.
The structural price is stated plainly in the paper: every block is
conditioned directly on x, and the authors note this “is different from a
standard deep neural network which does not have direct connections from the
input into each block”. That is what lets a block be trained without knowing
what the blocks below it computed — and it means the stack is not learning a
hierarchy of representations in the usual sense.
Three variants: discrete-time (NoProp-DT), continuous-time diffusion (NoProp-CT), and flow matching (NoProp-FM).
What was measured
MNIST, CIFAR-10 and CIFAR-100, no data augmentation, 15 values per cell (3 seeds × 5 inference runs). The backprop baseline was built to have the same forward structure, which makes it a fair comparison rather than a convenient one.
Test accuracy, discrete time:
| method | MNIST | CIFAR-10 | CIFAR-100 |
|---|---|---|---|
| Backprop (one-hot) | 99.46 | 79.92 | 45.85 |
| Backprop (prototype) | 99.44 | 79.58 | 47.80 |
| NoProp-DT (one-hot) | 99.47 | 79.25 | 45.93 |
| NoProp-DT (prototype) | 99.54 | 80.54 | 46.06 |
So NoProp-DT matches or slightly beats backprop on MNIST and CIFAR-10, and loses to it on CIFAR-100 by about 1.7 points.
Against prior backprop-free methods the margin is large: Forward-Forward 98.63 on MNIST; Local Greedy Forward Gradient 69.32 on CIFAR-10; Difference Target Propagation 50.71 on CIFAR-10, against NoProp-DT’s ~80.
Memory roughly halves. Discrete time: 0.49 / 0.64 / 1.23 GB against backprop’s 0.87 / 1.17 / 1.73. Continuous time the gap is much larger — NoProp-CT at 0.45–1.05 GB against the adjoint sensitivity method’s 2.32–6.45.
The number the table gives you that the text does not discuss
NoProp fits the training set much less well. CIFAR-10 train accuracy: backprop 99.98, NoProp-DT 95.02–97.23. CIFAR-100: backprop 98.63–99.19, NoProp-DT 83.25–90.70.
Matching test accuracy while being 8–15 points behind on train is a regularization-shaped result, not an optimization-parity one. On these datasets, where backprop is at ceiling on train and the test numbers are set by generalization, that costs nothing. It is exactly the situation that changes when the task is large enough that fitting the training data is the binding constraint — which is every setting this record’s practices are argued in. The paper does not raise this.
Conditions
MNIST and CIFAR, no augmentation. Three small image classification sets, and the largest model is whatever fits that.
The continuous-time variants are much weaker, and one fails. NoProp-CT reaches 33.66 on CIFAR-100 against NoProp-DT’s ~46; NoProp-FM with one-hot embeddings gets 6.38 ± 4.9 on CIFAR-100 — it does not learn. Learned embeddings rescue it (31.14). So “NoProp works” is carried by the discrete-time variant, and the headline should not be read as covering the family.
The comparison to other backprop-free methods is not like-for-like, and the authors say so: those methods “use different architectures and do not condition layers explicitly on image inputs as NoProp does — making direct comparisons challenging”. Giving every block the raw input is a large advantage and it is the thing being compared away.
The parallelism argument is motivational, not measured. The introduction names impeded parallel computation across devices as one of three reasons to want a backprop-free method. Nothing here measures a distributed run.
And the conclusion lists no limitations at all. It is a page of positive summary. The caveats above are recoverable from the tables and from one sentence in §2.1, not from a limitations section, because there is not one.