Learner · Results
Every number, with its label.
No backprop. No stored photos. Where a run differs from strict lessons, the label says so.
Vision.
| Measure | Result | Label |
|---|---|---|
| CIFAR-10, 5 eyes | 87.17% | ± 0.19, one look, three seeds, settings chosen on validation, nothing previewed. 5 eyes, each taught in a different lesson order: 5 passes, not one stream. |
| CIFAR-10, one strict eye | 84.93% | One look, three seeds. 86.00% with six looks. |
| CIFAR-10.1, fresh photos | 74.38% | The same 5 eyes on a separate photo set. |
| CIFAR-100, 7 eyes | 64.8% | One look, with an unlabelled preview. 65.46% with six looks. One ensemble. |
| CIFAR-100, one strict eye | 61.7% | 50 classes, then ten lessons of 5. One look, three seeds, no preview, no pretraining. |
| MNIST | 99.06% | Five looks, closed form, all at once, hand-set constants, one seed. |
| Fashion-MNIST | 91.73% | Two looks; 90.83% with one, same model. One seed. Best so far, pending reruns. |
Radar and speech.
| Measure | Result | Label |
|---|---|---|
| Radar, trained only on simulated images | 89.55% | ± 0.28, six looks, three seeds, tested on measured radar chips. With the vehicle blanked it scores 10.39%, near chance. DenseNet + noise: 77.7. Backprop ensembles: over 95. |
| KTH 77 GHz radar | 89.1% | Nine classes, whole recordings held out, three seeds. 98.1% drone or not. Dataset classification, not field detection. |
| Speech Commands, all 35 words | 91.2% | Average across seven lessons, three seeds, official test list. 88.9% final. |
Exact, not approximate.
| Measure | Result | Label |
|---|---|---|
| Delete 5,000 photos | 0.149 s | The sums match a model that never saw them, bit for bit. Answers match a clean run on 10,799 of 10,800. |
| Merge two copies | Bit-identical | Integer sums, shared eye, merged in opposite orders. |
| Readout live memory | 1.887 MB | Down from 7,835 MB, and 612 s down to 11.10 s, while CIFAR-10 at six looks rose from 83.10% to 85.21%. Readout stage only; eye and on-disk codes counted separately. |
| Five-dataset continual text | 71.02% | 575,000 documents across five sources, zero replay, no backprop. Unsupervised source router: 94.18% at 1.54M tokens/s in 32 MB RAM. |
Timing and small whole models.
| Measure | Result | Label |
|---|---|---|
| Small whole model, MNIST | 97.40% | 150.46 KB whole model (89.39% on Fashion-MNIST), no backprop, no matrix solve. Timing-based eye, local error rule on one layer, multi-pass, all at once, linear reader at test, one seed. |
| Timing-neuron sequences | 3,963 | Named at 99% with 16 timing neurons over three seeds on synthetic temporal sequences, read by template match over real neuron traces. |
| Graded-release firing threshold | 99.57% | ± 0.44, three seeds, decided by a true firing threshold on synthetic symbols, cutting stray-noise false fires from 27.0% to 0.50%. |
| Two-beat chords with check neurons | 97.92% | 95.75% across 2 slots and 91.58% across 4 slots over three seeds. |
Published reference points.
| Method | CIFAR-10 | Setting |
|---|---|---|
| Myrtle kernel | 90% | All classes at once; fixed kernel. |
| SimplePatch | 88.5% | All classes at once; classifier on top trained by SGD. |
| This learner, 5 eyes | 87.17% | Five lessons of two classes; no backprop; no stored photos. |
| Random patches | 85.6% | All classes at once. |
| SoftHebb (ICLR 2023) | 80.3% | All classes at once; Hebbian hidden layers, then a supervised head. |
| RanDumb (NeurIPS 2024) | 55.6% | Five lessons of two classes; raw-pixel random features; no exemplars; online. |
Protocols are not identical, and these are our own runs, not yet independently replicated. An earlier 86.0% headline used CIFAR-10 test subsets to choose settings and has been retired. Bring a harder test: you keep the labels, we return predictions.