Models / Dilated CNN / Workspace
code

Dilated CNN · Live Inference

Real forward + backward passes on CPU: predictions, Grad-CAM, and robustness probes.

Dilated CNN · CPU
analytics88.4% AccCIFAR-102.4M paramsIn this workspace:Grad-CAMFGSM attackOOD signal
Recognizes10 everyday object categories10 classes · airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck
Accuracy88.4% Acc88.2% on the CIFAR-10 test set
Works best onA photo of one of the 10 listed objectsDownscaled to 32×32 color — fine detail is thrown away.
warning
Expect confident mistakes outside its scope

It recognizes only these 10 categories, and only at a tiny 32×32 resolution. Anything else, a person, a building, a banana, is forced into the nearest of the ten, confidently and wrongly. Even valid objects can be misread once detail is lost at 32×32, so treat errors as part of the demo.

Concepts in play
view_in_arReceptive FieldsThe patch of the input a neuron can see — tiny at first, growing with depth, stride, and dilation.filter_center_focusConvolutionOne small grid of weights, slid across every position, turns pixels into feature maps by dot product.

Top-5 Predictions

Upload an image and run analysis to see predictions.

Input

image

Activation Map

LOW→HIGH
blur_on

Layer Source

—

Focus Metric

—

Latency

—

Confidence (top-1)

—

Run analysis to compute

Inference Latency

—

CPU forward pass