Unverified50% confidenceFactExact time
The third model uses the Arabic Digits MNIST dataset, with an architecture of 784→20→20→15→10, cross-entropy loss plus plain gradient descent, an accuracy of about 96% after 2000 epochs, and uses JAX automatic differentiation to compute gradients.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe first model uses the original MNIST dataset, with an architecture of 784→20→20→10, MSE as the loss function, plain gradient descent as the optimizer, no activation function in the output layer, and a test accuracy of about 92% after 2000 epochs of training.72% similarUnverifiedThe second model's architecture is expanded to 784→40→40→10, MSE is replaced with cross-entropy, softmax is added to the output layer, the optimizer remains plain gradient descent, and the accuracy is about 96% after 2000 epochs.68% similarVerified4-bit quantization reduces model size to 1/8 of the original at the cost of approximately 1–3% accuracy loss.63% similarUnverifiedIn ablation experiments, replacing numerical EHR encoding with text serialization resulted in a 3.66 percentage point drop in AUROC, the largest impact among all interventions tested.62% similarUnverifiedIn the fixed-vector self-similarity experiment, when the offset goes from 0 to 1024, the mean of the raw scores always fluctuates within ±0.38, while the mean of the absolute values only drops from 3.17 to 3.02, a change of about 4%.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1126396API
curl https://kongchang.com/api/v1/knowledge/claims/1126396MCP
get_claim(id=1126396)