Deep Learning Lab

Deep Learning Training Lab

Deep Learning stacks many layers — enabling the model to learn hierarchical representations from raw features: from edges → shapes → patterns → diagnosis.

Adjust architecture and watch the loss curve descend as the model trains epoch by epoch.

Batch Gradient Descent

The network learns by iteratively adjusting all weights to minimise the loss. Each epoch is one full pass through the training set:

For each epoch: 1. Forward pass → compute ŷ 2. Compute Loss (BCE) 3. Backprop → compute ∂L/∂w 4. Update: w ← w − η·∂L/∂w Loss should decrease monotonically (if learning rate is appropriate)

Batch Size & Dropout

Batch size controls how many patients are used per weight update. Smaller batches → noisy but faster convergence. Larger → smooth but slower.

Dropout randomly silences neurons during training — prevents the network from memorising individual patient noise.

Dropout(p=0.5): Each neuron disabled independently with probability p during training. → Forces redundant representations → Acts as ensemble of sub-networks

⚖️ Ethical Reality

Deep networks trained on ImageNet-like medical imaging datasets often perform worse on patients from hospitals with different scanner hardware. The model learns the scanner's artefacts, not the pathology. External validation across institutions is mandatory before clinical deployment.

① Architecture

Hidden Layers4

depth of the network

Neurons per Layer32

width of each hidden layer

Dropout Rate0.3

regularisation (0 = none)

② Training Parameters

Learning Rate (η)0.010

too high = diverge · too low = slow

Batch Size32

patients per gradient update

Epochs50

training iterations

③ Training Metrics

Final Train Accuracy

—

Loss: —

—

Parameters

—

Epoch

—

Val Accuracy

—

Overfit Gap

④ Training Curve — train loss (violet) · validation loss (cyan) · accuracy (green) · overfit gap = generalisation failure

⑤ ANN Architecture — signals flow left→right (forward pass, violet) then right→left (backprop, cyan) · faded nodes = dropout silenced · brighter = higher activation