Borrow. Adapt.
See further.

Bring your knowledge of convolution, pooling, gradient descent, and train/validation/test splits. We write a network as a backbone f_θ followed by a task head g_φ. θ and φ are learned parameters.

Split the 120 original images first, preferably by plant or capture session if images are related. For illustration: 84 train, 18 validation, 18 test, stratified if feasible. Use validation to select a recipe and touch the test set only for the final evaluation. A tiny split has high uncertainty; repeated or cross-validation can help.

The labs below use exact small numerical examples, hand-designed features, or architecture counts. No accuracy curves or learned embeddings are claimed.

Early CNN layers often detect simpler patterns and later ones are more task-specific, but this is a useful tendency, not a guarantee. Source bias and domain shift can make a pretrained representation unhelpful (negative transfer).

ImageNet ResNet-18 produces 512 features after global average pooling. Its original head has 512 × 1,000 + 1,000 = 513,000 parameters. A 3-class replacement has 512 × 3 + 3 = 1,539. Pretraining changes the initialization; whether the backbone later changes is a separate decision.

Use the checkpoint's expected channel order, input range, resizing, and normalization. The weights alone are not the complete input contract. Training augmentation may vary the crop, but normalization must remain compatible.

Decision: compare a frozen-backbone baseline against selective fine-tuning on validation data. Do not promise a fixed accuracy gain or claim that more pretraining always wins.

For channel 1, [[1,3],[2,2]] averages to 2. Channel 2, [[0,2],[4,2]], also averages to 2. Thus z = [2,2]. Let W = [[1,−0.5],[−0.5,1]] and b = [1,1]. The scores are [2,2] and softmax gives [0.5,0.5].

For a labelled example y, cross-entropy is L = −log p_y. Optimize W and b while keeping θ fixed. With z = [3,1], the scores become [3.5,0.5], so p_A ≈ 0.953. This confidence comes from a hand-selected toy classifier; it is not calibrated evidence about real leaves.

Feature maps are generally learned channel responses, not guaranteed human-interpretable properties. The channels here are deliberately simple.

Cached features can save repeated backbone computation when the backbone and preprocessing are fixed. Caching one view per image removes fresh online augmentation; cached augmented views or recomputation are needed for varied views.

Forward pass: frozen layers still transform x. Backward pass: frozen parameters need no parameter gradients; gradients through their operations may still be needed if earlier trainable layers require them. Do not indiscriminately detach tensors in a partially trainable network.

In PyTorch, requires_grad_(False) controls parameter gradients; eval() controls module behaviour. They are different. BatchNorm running means and variances are buffers and can change in training mode even when affine weights are frozen. Dropout also behaves differently in train and eval modes.

For a fixed-feature baseline: freeze the backbone, put it in eval mode, train the head, and optimize only parameters that require gradients. If a parent train() call resets the backbone mode, restore its eval mode. When unfreezing, update the optimizer's parameter groups and choose the BatchNorm policy deliberately.

Counts below are for torchvision-style ResNet-18: stem + layer1 = 157,504; layer2–4 = 11,019,008; a new 3-class head = 1,539. The grouped blocks simplify the decision; they are not individual layers.

For width W, a horizontal flip is x′[r,c] = x[r,W−1−c]. In the toy grid, brightness adds δ and clips: x′ = min(9,max(0,x+δ)). A right shift inserts zeros on the left and discards the rightmost column. A clockwise rotation maps output [r,c] to input [H−1−c,r] for a square grid.

Split original examples before generating augmented views. Put related captures in the same split when necessary. Otherwise a transformed training image can leak into validation and make performance look better than it is.

Use random augmentation for training and a consistent evaluation preprocessing pipeline. Deliberate test-time augmentation is a separate technique, not the default validation recipe. A severe crop can remove the object; colour shifts can erase a disease cue. Stronger is not automatically better.

For leaf species, moderate position and illumination changes may preserve labels. For left/right arrow classes, a horizontal flip requires changing the label or excluding that transform. For colour-based diagnoses, brightness/colour changes need domain care.

For L stride-1 3 × 3 convolutions without dilation, r = 1 + 2L. Three layers give 7 × 7 theoretical coverage. This is a geometric upper bound; the effective influence of individual pixels need not be uniform. Deeper layers, dilation, pooling, global pooling, large kernels, and hybrid attention can all widen context.

A conventional convolution uses the same learned spatial weights at different locations; attention weights depend on the current token content. Standard ViT splits an image into patches, flattens and projects each patch to an embedding, adds positional embeddings, then uses Transformer blocks. The original classification architecture also has a class token.

For 224 × 224 images and 16 × 16 patches, N = 196 patch tokens. There are 196² = 38,416 patch-to-patch attention scores per head. Including the class token gives 197² = 38,809. These counts are not total FLOPs or whole-model memory. Doubling patch width to 32 gives 49 tokens and 2,401 patch pairs, but coarser spatial detail.

Full attention has quadratic token-pair cost and weaker built-in locality than a CNN. Large-scale pretraining and the training recipe matter; ViTs are not automatically better on small datasets. The motivation is a different way to connect information, not a claim that CNNs are incapable of global reasoning.