Artificial neuron calculator
Compute z=w₁x₁+w₂x₂+b, then apply sigmoid.
Interactive bench
Change inputs. Watch the mechanism.
Every plot has labels and a text readout.
Activation oscilloscope
Text equivalent: selected function, current output and derivative are shown above.
DNN parameter counter
Enter layer widths, e.g. 4, 3, 2. Each dense layer has nᵢnᵢ₊₁+nᵢ₊₁ parameters.
Perceptron: one update
Rule: w←w+η(y−ŷ)x; b←b+η(y−ŷ).
Gradient descent trajectory
Objective J(θ)=θ², gradient 2θ.
Backprop chain-rule tracer
For ŷ=wx+b; L=½(ŷ−y)².
Optimizer stepper
Same gradient sequence; compare adaptive step sizes.
AdagradRMSPropAdadelta
| Optimizer | θ now | Memory | Interpretation |
|---|
Dropout visualizer
Bias–variance curve explorer
Complexity moves model from underfit → useful fit → overfit.
training errorvalidation error
Mechanism maps
Where information moves
Rapid recall
Formula + comparison sheet
One screen before the exam.
| Compare | First | Second | Exam-safe distinction |
|---|---|---|---|
| Parameter / hyperparameter | W, b; learned | η, batch size, depth; chosen | Model learns parameters; validation helps choose hyperparameters. |
| Batch / SGD / mini-batch | All samples / one sample | Small subset | Mini-batch balances stable vectorized updates and useful noise. |
| Underfit / overfit | Train & validation errors high | Train low, validation high | High bias versus high variance. |
| L1 / L2 | Sparse, absolute penalty | Smooth shrinkage, squared penalty | L1 can set weights exactly zero; L2 usually does not. |
| Backprop / optimizer | Computes gradients | Uses gradients to update | Backprop is not an optimizer. |
| Neural bias / statistical bias | Trainable offset b | Systematic modeling error | Same word, different concepts. |
Test mode
10-question mid-term drill
Score is saved locally.