ResNet-18 · Live Inference
Real forward + backward passes on CPU: predictions, Grad-CAM, and robustness probes.
analytics76.6% AccCIFAR-10011M paramsIn this workspace:Grad-CAMCorruptionsOOD signal
Recognizes100 fine-grained categories100 classes · 100 CIFAR-100 categories (apple, bicycle, dolphin, maple tree, …)
Accuracy76.6% Acc76.6% on the CIFAR-100 test set
Works best onA photo of one of the 100 CIFAR-100 categoriesDownscaled to 32×32 color.
warning
Expect confident mistakes outside its scope
100 categories at 32×32 is genuinely hard: even on in-scope images this model is right roughly 3 times in 4, and confusions between look-alike classes (leopard vs. tiger, boy vs. girl) are normal. Inputs outside the 100 classes will simply be mislabeled.
Concepts in play
trending_downVanishing GradientsBackprop multiplies gradients layer by layer; in deep nets they shrink toward zero unless a skip connection carries them through.filter_center_focusConvolutionOne small grid of weights, slid across every position, turns pixels into feature maps by dot product.view_in_arReceptive FieldsThe patch of the input a neuron can see — tiny at first, growing with depth, stride, and dilation.
Top-5 Predictions
Upload an image and run analysis to see predictions.
Input
Activation Map
LOW→HIGH
Layer Source
—
Focus Metric
—
Latency
—
Confidence (top-1)
—
Run analysis to compute
Inference Latency
—
CPU forward pass