ResNet-50 · Live Inference
Real forward + backward passes on CPU: predictions, Grad-CAM, channel attribution, and robustness probes.
analytics76.01% top-1ImageNet-1k25.6M paramsIn this workspace:Grad-CAMChannel attributionFGSM attackCorruptionsOOD signal
Recognizes1000 ImageNet object categories1000 classes · 1000 ImageNet-1k labels — common objects, animals, foods, and scenes
Accuracy76.01% top-176.0% top-1 on ImageNet-1k
Works best onA clear photo of a common object, animal, or sceneResized to 224×224 color.
info
Closed-set model — keep in mind
The most capable model here, but still closed-set: it must choose one of 1000 ImageNet labels. Screenshots, diagrams, text, or unusual crops fall outside that vocabulary and get forced into the nearest label anyway.
Concepts in play
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.trending_downVanishing GradientsBackprop multiplies gradients layer by layer; in deep nets they shrink toward zero unless a skip connection carries them through.
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