Experimental learning
Learning, one example at a time.
JAYCE explores Adaptive Prototype Memory (APM): learning from labeled examples and corrections to improve future predictions.
The name stands for Jayce Associates Your Categorized Exemplars, inspired by a toddler learning toy names.
How APM differs from backpropagation
A prototype starts as the numerical representation of a labeled example, or exemplar. APM learns by saving prototypes or moving them toward new examples.
| Step | APM | Neural network with backpropagation |
|---|---|---|
| Stores | Labeled prototypes. | Weights connecting neurons. |
| Learns | Saves or moves a prototype. | Uses error gradients to update weights. |
| Predicts | Uses the label of the closest prototype. | Passes the input through the trained network. |
In the Java benchmarks, the neural network had higher accuracy in 8 of 12 comparisons; APM led on the digit tests. The results also compare training time and memory use.
Learning in the chat app
Jayce tries → the parent answers → APM learns → Jayce tries again.
A parent model such as Qwen answers every turn. When its answer differs from Jayce's, Jayce learns from it, updates its prototype memory, and tries again. The parent's weights stay fixed. Jayce can learn its mistakes too.
Watch Jayce learn toy names
Step through new toys, corrections, and the predictions that follow.
Jayce's toy bins
6 labels · 2 memory slots eachHow Jayce compares examples
Toy features, simplifiedCheck the distances Square each difference, then add them.
Why are toys close?
The map uses illustrative body-length and bulk scores. Similar scores place toys nearby, even when their labels differ.
What does the correction do?
A person supplies the correct label. Jayce saves the example in that bin, or moves its closest prototype if both slots are full.