Neural Networks and Deep Learning

Introduction and Context

Neural networks are function-approximation systems composed of interconnected computational units. Their modern importance arises from the ability to learn hierarchical representations directly from data: early layers may detect simple patterns, while later layers combine them into increasingly abstract concepts. Deep learning refers principally to the use of networks with many trainable layers, together with large datasets, effective optimization methods, specialized hardware, and representation-learning objectives.

Neural networks are not universally superior to classical machine-learning methods. On small, structured tabular datasets, carefully tuned tree ensembles may outperform deep models. Deep learning is especially valuable when the input contains complex structure—images, speech, language, sequences, graphs, or multimodal observations—and when sufficient data or pretrained representations are available.

Learning Objectives

After studying this chapter, a learner should be able to:

1. Explain multilayer perceptrons, forward propagation, loss functions, and backpropagation.
2. Compare activation functions and identify their optimization consequences.
3. Analyze gradient-based optimization, initialization, normalization, and regularization.
4. Select convolutional, recurrent, attention-based, and transformer architectures for different data types.
5. Design an appropriate neural-network workflow for tabular, image, text, and multimodal problems.
6. Evaluate reliability, fairness, interpretability, computational cost, and deployment risks.

1. The Multilayer Perceptron

A perceptron computes a weighted sum of inputs followed by a threshold or nonlinear transformation. A multilayer perceptron (MLP) stacks affine transformations and nonlinear activation functions. For layer ℓ,

Here, a^(ℓ−1) is the input activation, W^(ℓ) is a weight matrix, b^(ℓ) is a bias vector, z^(ℓ) is the pre-activation, and φ^(ℓ) is applied elementwise. The input is a^(0) = x. The final activation is selected according to the prediction task.

For binary classification, a common output is the sigmoid function:

It maps a real-valued logit to a number between 0 and 1. For K-class classification, softmax gives

where p_k is the predicted probability of class k. Softmax probabilities sum to one. For numerical stability, implementations subtract the largest logit before exponentiation; this leaves the probabilities unchanged.

For regression, the output layer is usually linear, ŷ = z, although transformations may be used when the target has constraints such as positivity.

Expressive Power

A network with at least one hidden layer and a suitable nonconstant activation can approximate broad classes of continuous functions on compact domains. This universal-approximation result does not imply that the network will learn the function efficiently, generalize well, or require only one hidden layer in practice. Depth often represents compositional structure more efficiently than a shallow network.

2. Activation Functions

Activation functions introduce nonlinearity. Without them, a composition of affine transformations remains affine, regardless of the number of layers.

The derivative of sigmoid is σ′(z)=σ(z)(1−σ(z)). Its maximum is 1/4, so repeated multiplication through many layers can make gradients very small. ReLU has derivative 1 for positive inputs and 0 for negative inputs (apart from the undefined point at zero, handled by an implementation convention). This helps optimization but can create dead neurons when parameters place most inputs in the negative region.

Activation choice is connected to initialization, normalization, architecture, and task. The output activation must also match the likelihood or target representation: sigmoid for independent binary outcomes, softmax for one mutually exclusive class among K classes, and often a linear output for unconstrained regression.

3. Loss Functions and Probabilistic Interpretation

Training minimizes an empirical objective. Given n examples, a regularized objective can be written as

where θ contains all parameters, f_θ is the network, ℓ is the per-example loss, λ≥0 controls regularization, and Ω penalizes selected parameters.

For binary classification with label y∈{0,1} and predicted probability p, binary cross-entropy is

For one-hot multiclass targets, categorical cross-entropy is

Cross-entropy is the negative log-likelihood under Bernoulli or categorical assumptions. With a sigmoid or softmax output, it strongly penalizes confident incorrect predictions. For regression, mean squared error is

It corresponds to a Gaussian-noise likelihood under a constant-variance assumption. Mean absolute error is more robust to some outliers but has a nondifferentiable point at zero, normally managed with subgradients or a smooth approximation.

4. Backpropagation

Backpropagation efficiently computes derivatives of the loss with respect to every parameter by applying the chain rule from the output layer toward the input. It is not itself an optimization algorithm; it supplies gradients to algorithms such as stochastic gradient descent.

Consider a single example and a two-layer network:

If the loss is ℓ(ŷ,y), define δ^(ℓ)=∂ℓ÷∂z^(ℓ). The output-layer error is

where ⊙ denotes elementwise multiplication. For a hidden layer,

The parameter gradients are

For a minibatch, these quantities are summed or averaged according to the implementation. Matrix dimensions must agree: if a layer has m output units and q input units, W has shape m × q, b has shape m, and the gradient with respect to W has the same shape as W.

For softmax combined with categorical cross-entropy, the derivative with respect to the logits simplifies to

This simplification is both mathematically useful and numerically more stable than separately differentiating potentially extreme exponential values.

5. Optimization and Training Dynamics

The basic gradient-descent update is

where η>0 is the learning rate and ∇J is the gradient. Batch gradient descent uses all training examples; stochastic gradient descent uses one; minibatch training uses a subset and is the dominant practical approach.

Momentum accumulates a direction over time. Adaptive methods such as Adam maintain estimates of first and second moments of gradients. Their effectiveness does not remove the need for learning-rate selection, validation, and stability checks. A large learning rate can cause divergence; a small one can make training impractically slow or trap optimization in poor regions.

Important training controls include:

• Learning-rate schedules: step decay, cosine decay, warm-up, or validation-based reduction.
• Batch size: affects computational efficiency, gradient noise, and sometimes generalization.
• Early stopping: stops training when validation performance ceases to improve.
• Gradient clipping: limits unusually large gradients, especially in recurrent networks.
• Checkpointing: saves parameters associated with the best validation criterion rather than merely the final iteration.

Initialization

Poor initialization can cause exploding or vanishing activations and gradients. Xavier/Glorot initialization is commonly suited to approximately symmetric activations; He initialization is designed for ReLU-like activations and uses a variance related to the number of incoming units. Exact distributions and scaling depend on fan-in, fan-out, activation, and framework convention. Biases are often initialized at zero, although specialized architectures may use different choices.

Normalization

Batch normalization standardizes intermediate activations using minibatch statistics during training and moving estimates during inference. Layer normalization normalizes across features within an individual example and is particularly compatible with sequence models and transformers. Normalization can improve conditioning, but it changes the optimization and inference behavior; it is not a substitute for sound data preprocessing.

6. Regularization and Generalization

A model generalizes when it performs well on unseen data drawn from the intended population. Deep networks can have many more parameters than training examples yet still generalize, because optimization bias, data structure, augmentation, and regularization constrain the functions actually learned.

Common methods include:

1. Weight decay: adds a penalty proportional to the squared parameter norm. Under the conventional objective, Ω(θ)=½||θ||²₂. It discourages large weights, although decoupled weight decay in optimizers such as AdamW is not identical to adding an L2 penalty in every setting.
2. Dropout: during training, randomly masks units and rescales retained activations. It reduces reliance on particular paths but may be unnecessary or harmful in some modern architectures.
3. Data augmentation: creates label-preserving variations, such as crops for images or controlled perturbations for audio. Augmentations must preserve the problem’s semantics.
4. Label smoothing: replaces hard target probabilities with a less extreme distribution, reducing overconfidence in classification.
5. Pretraining and transfer learning: initialize a model from a broader dataset, then adapt it to a target task. This can improve accuracy and reduce data requirements, but distribution mismatch can cause negative transfer.

Data splitting must occur before operations that could leak information. For example, normalization statistics, feature selection, augmentation policy design, and pretrained adaptation must not use held-out test information. In temporal or grouped data, random splitting may be invalid; chronological or group-aware splitting is often required.

7. Convolutional Neural Networks

Convolutional neural networks (CNNs) exploit local connectivity and shared parameters. A two-dimensional cross-correlation, commonly called convolution in deep-learning software, is

where X is an input image or feature map and K is a kernel. Stride controls the displacement between applications; padding controls boundary treatment; multiple kernels produce multiple output channels. For a one-dimensional input of length L, kernel size k, padding p, and stride s, the output length is

assuming dilation 1 and integer floor operation.

CNNs are effective for images because nearby pixels often have related meaning and the same visual feature may occur at different locations. Pooling or strided convolution reduces spatial resolution, increasing receptive field and reducing computation. Residual connections learn a correction to an identity mapping, helping optimization in deep networks.

CNN limitations include sensitivity to distribution shifts, possible loss of fine spatial information, and substantial data and compute requirements. Vision transformers can replace convolutions by processing image patches with attention, although hybrid designs remain common.

8. Recurrent Networks and Sequential Data

A recurrent neural network maintains a state that evolves over time:

where x_t is the input at time t and h_t is the hidden state. The same parameters are reused across positions, allowing variable-length sequences. Backpropagation through time unfolds the recurrence, but repeated Jacobian multiplication can create vanishing or exploding gradients.

Long short-term memory (LSTM) and gated recurrent unit (GRU) architectures use gates to regulate information flow. LSTM gates include input, forget, and output controls and a cell state. They can model longer dependencies than a basic RNN, but recurrent computation is sequential and therefore less parallelizable than attention.

RNNs remain useful for streaming, low-latency, or resource-constrained tasks. For long-range language modeling and large-scale training, attention and transformers have largely become dominant.

9. Attention and Transformers

Attention allows a representation at one position to selectively combine information from other positions. Scaled dot-product attention is

Q, K, and V are query, key, and value matrices; d_k is the key-vector dimension. Scaling by √d_k prevents dot products from becoming excessively large, which would make softmax gradients small. Multi-head attention projects inputs into several subspaces, performs attention independently, and concatenates the results.

A transformer layer generally contains multi-head attention, a position-wise feed-forward network, residual connections, and normalization. Because attention alone does not encode order, positional information must be supplied through learned positional embeddings, fixed functions, relative positions, or rotary representations.

In an autoregressive decoder, a causal mask prevents position t from attending to future positions. Encoder-only models are well suited to representation and classification tasks; decoder-only models generate sequences; encoder–decoder models map one sequence or modality to another.

Attention has a conventional quadratic cost in sequence length for dense self-attention: approximately proportional to n²d, where n is sequence length and d is representation width. This motivates sparse, linearized, recurrent, retrieval-augmented, and long-context attention methods. Longer context also increases memory, latency, and opportunities for irrelevant or misleading information.

10. Architecture Selection Across Data Modalities

Tabular Data

An MLP can process normalized numerical features and embeddings for categorical variables. Missingness indicators, careful target encoding, and leakage-resistant validation are often more important than increasing depth. Tree-based baselines should be included. Neural additive models, tabular transformers, and entity embeddings can be useful when there are many categories, relational structure, or large-scale data.

Images

CNNs offer strong inductive bias and efficient local processing. Transfer learning from a pretrained vision model is usually preferable to training from scratch unless the dataset is large or the domain is highly specialized. Evaluation should consider class imbalance, calibration, subgroup performance, and whether image acquisition conditions differ between training and deployment.

Text

Tokenization converts text into discrete units, which are mapped to vectors. Transformer encoders support classification and retrieval; decoder models support generation; encoder–decoder models support translation and structured transformation. Fine-tuning, parameter-efficient adaptation, retrieval augmentation, and prompt-based methods should be compared against simpler lexical or linear baselines.

Multimodal Data

A multimodal system may use early fusion (joining features), late fusion (combining predictions), or intermediate fusion (cross-attention between modality representations). Missing modalities, asynchronous timestamps, inconsistent sampling, and spurious correlations are central design concerns. Modality-specific encoders followed by a shared projection space can support contrastive alignment, but similarity in embedding space does not guarantee factual or causal alignment.

11. Worked Example: A Small Classifier

Suppose a binary classifier produces logit z=2. Its predicted probability is

p=1÷(1+e^(−2))≈0.881.

If the true label is y = 1, the binary cross-entropy is

−log(0.881)≈0.127.

For sigmoid output combined with binary cross-entropy, the derivative with respect to the logit is p−y≈−0.119. Thus, gradient descent will tend to increase the logit for this correctly classified but not perfectly confident positive example. If instead y = 0, the loss would be approximately −log(1−0.881)≈2.127, and the derivative would be approximately 0.881, producing a strong update that lowers the logit.

The example illustrates why cross-entropy distinguishes confident correct predictions from confident errors more sharply than an uncalibrated accuracy measure.

12. Applied Case: Illustrative Hospital Imaging System

Background. A hospital seeks to identify pulmonary abnormalities in chest radiographs. The proposed system is illustrative, not a report of a specific institution.

Problem. The dataset contains images from several hospitals, labels produced partly by radiologists and partly by automated report mining, and substantial class imbalance. Images from the same patient may occur more than once.

Analysis. A random image-level split would leak patient-specific patterns and overestimate generalization. Differences in portable versus fixed imaging devices may act as shortcuts. A high area under the ROC curve would not by itself establish clinical usefulness, because threshold selection, calibration, sensitivity at an operational specificity, and subgroup performance matter.

Approach. The team uses patient-level and hospital-aware splits, retains a locked test set, and compares a pretrained CNN with a simpler baseline. It audits labels, records acquisition metadata, uses class-weighted or sampled training where justified, and evaluates calibration and decision curves. A model card documents intended use, contraindications, training population, uncertainty, and known failure modes. The system is positioned as decision support rather than autonomous diagnosis. Human reviewers can override it, and monitoring detects drift in devices, patient populations, and prevalence.

Outcome. The CNN improves sensitivity at the selected operating point, but performance decreases on an external hospital. Calibration is repaired using validation data from the deployment context, and the deployment scope is narrowed until prospective monitoring provides stronger evidence.

Learning points. Architecture is only one part of a reliable system. Dataset provenance, leakage control, external validation, calibration, workflow integration, accountability, and post-deployment monitoring are equally important.

13. Evaluation, Interpretability, and Governance

Accuracy can be misleading for imbalanced tasks. Precision, recall, specificity, F1 score, area under the precision–recall curve, calibration error, and task-specific utility should be selected according to consequences. For regression, MAE, RMSE, calibration of predictive intervals, and subgroup error are useful alternatives.

Interpretability methods include saliency maps, feature occlusion, integrated gradients, attention visualization, counterfactual analysis, and surrogate models. These methods answer different questions and can be unstable or misleading. Attention weights should not automatically be treated as explanations of causation.

Major risks include privacy leakage, memorization, copyright and data-licensing violations, demographic bias, automation bias, adversarial manipulation, insecure model endpoints, and environmental cost. Governance should specify data rights, access controls, retention, documentation, auditability, human oversight, incident response, and criteria for withdrawal. In regulated domains, model development must be aligned with applicable safety, privacy, medical, financial, or employment requirements.

14. Contemporary Directions and Best Practices

Current research includes self-supervised pretraining, foundation models, parameter-efficient fine-tuning, retrieval-augmented generation, diffusion models, mixture-of-experts systems, neural operators, graph neural networks, and efficient on-device inference. These developments increase capability but also intensify concerns about evaluation contamination, reproducibility, compute concentration, and deployment risk.

A disciplined project workflow is:

1. Define the decision, users, constraints, and failure costs.
2. Establish leakage-resistant splits and strong non-neural baselines.
3. Inspect data quality, missingness, imbalance, provenance, and subgroup coverage.
4. Select an architecture whose inductive bias matches the modality.
5. Train with reproducible seeds, tracked configurations, validation-based selection, and checkpointing.
6. Evaluate discrimination, calibration, robustness, fairness, latency, memory, and cost.
7. Perform external or temporal validation before operational use.
8. Document limitations and monitor drift after deployment.

Chapter Summary

Neural networks learn compositions of affine transformations and nonlinear activations. Multilayer perceptrons provide a general foundation, while convolutional networks exploit spatial locality, recurrent networks represent sequential state, and attention-based transformers model flexible relationships among positions or modalities. Backpropagation computes gradients through the chain rule; optimizers use those gradients to update parameters. Activation functions, initialization, normalization, learning-rate schedules, and regularization strongly influence training and generalization. Cross-entropy and squared-error losses have probabilistic interpretations under specific assumptions. Architecture selection must reflect data structure, computational constraints, and deployment requirements. Reliable deep learning additionally requires leakage-resistant evaluation, calibration, fairness analysis, security, documentation, human oversight, and monitoring.

Key Terms

Multilayer perceptron: A feed-forward network with one or more hidden layers.

Activation function: A nonlinear function applied to a layer’s pre-activation.

Logit: An unconstrained score before a sigmoid or softmax transformation.

Softmax: A function converting class scores into probabilities that sum to one.

Cross-entropy: Negative log-likelihood loss for probabilistic classification.

Backpropagation: Efficient reverse-mode computation of parameter gradients.

Gradient descent: Optimization by moving opposite the loss gradient.

Learning rate: Step-size parameter controlling each optimization update.

Weight decay: A regularization mechanism discouraging large parameters.

Dropout: Stochastic masking of units during training.

Batch normalization: Normalization using minibatch statistics during training.

Convolution: Local, shared-weight operation producing feature maps.

Receptive field: Input region capable of influencing a unit.

Recurrent neural network: A network whose state depends on prior sequence states.

LSTM: A gated recurrent architecture designed to preserve information over time.

Attention: Weighted combination of representations selected by queries and keys.

Transformer: An architecture built primarily from attention and feed-forward blocks.

Transfer learning: Reusing learned representations for a new task.

Calibration: Agreement between predicted probabilities and observed frequencies.

Data leakage: Unintended use of information unavailable at prediction time.

Review Questions

Short Answer Questions

1. Why are nonlinear activation functions necessary in a multilayer network?
2. Distinguish backpropagation from gradient descent.
3. State two causes of vanishing gradients and one mitigation strategy.
4. Why does scaled dot-product attention divide by √d_k?
5. Give one reason patient-level splitting is necessary in medical-image evaluation.

Long Answer Questions

1. Derive the hidden-layer backpropagation equation and explain the dimensions of each term.
2. Compare sigmoid, tanh, ReLU, and GELU with respect to gradients and typical use.
3. Explain how regularization, augmentation, early stopping, and transfer learning affect generalization.
4. Compare CNNs, RNNs, and transformers for image, streaming sequence, and long-context language tasks.
5. Design a complete evaluation plan for a multimodal prediction system, including leakage control and governance.

Analytical / Application Questions

1. A classifier has excellent accuracy but poor minority-class recall and unreliable probabilities. Which metrics, training adjustments, and deployment safeguards should be considered?
2. A transformer performs well on randomly split transaction records but poorly on future months. Diagnose likely causes and propose a revised experimental design.
3. Design an architecture for combining medical images and clinical notes. Justify the fusion strategy, missing-modality handling, validation split, and human-oversight process.

Multiple Choice Questions with Answers

Question 1

Which expression gives the gradient of sigmoid cross-entropy with respect to a binary logit?

A. p + y

B. p − y

C. y − 1

D. p(1 − p)

Correct Answer: B. p − y

Question 2

What is the primary purpose of nonlinear activations in a multilayer network?

A. To make all outputs probabilities

B. To eliminate the need for data

C. To prevent a composition of layers from collapsing to one affine transformation

D. To guarantee global optimization

Correct Answer: C. To prevent a composition of layers from collapsing to one affine transformation

Question 3

For a one-dimensional input of length 10, kernel size 3, padding 1, and stride 2, what is the output length?

A. 4

B. 5

C. 6

D. 7

Correct Answer: B. 5

Question 4

Which architecture most directly exploits local spatial connectivity and shared visual filters?

A. Convolutional neural network

B. Basic recurrent neural network

C. Linear regression

D. Naive Bayes classifier

Correct Answer: A. Convolutional neural network

Question 5

Why is √d_k used in scaled dot-product attention?

A. To force attention weights to be uniform

B. To reduce the magnitude of dot products and improve softmax conditioning

C. To remove positional information

D. To make sequence length constant

Correct Answer: B. To reduce the magnitude of dot products and improve softmax conditioning

Question 6

Which practice is most appropriate for preventing patient-level leakage in an imaging dataset?

A. Randomly split individual image files

B. Use the test set to choose the threshold

C. Split by patient before model selection

D. Duplicate rare images across all splits

Correct Answer: C. Split by patient before model selection

Question 7

Which output activation is generally appropriate for mutually exclusive K-class classification?

A. Linear

B. ReLU

C. Softmax

D. tanh

Correct Answer: C. Softmax

Question 8

What is a defining computational limitation of dense self-attention for sequence length n?

A. It cannot process variable-length inputs

B. Its pairwise attention computation grows approximately quadratically with n

C. It requires a convolutional kernel

D. It cannot use positional information

Correct Answer: B. Its pairwise attention computation grows approximately quadratically with n

Question 9

Which method directly reduces reliance on particular hidden-unit paths during training?

A. Dropout

B. Label encoding

C. One-hot output coding

D. Min–max target definition

Correct Answer: A. Dropout

Question 10

A model predicts p ≈ 0.881 for a positive example. What is its binary cross-entropy approximately?

A. 0.013

B. 0.119

C. 0.881

D. 2.127

Correct Answer: B. 0.119