Unit 4: Evolution and Modern AI Paradigms - Subjective Questions

Symbolic Artificial Intelligence is an approach in which knowledge is represented explicitly using symbols, facts, rules, and logical relationships. Reasoning is performed by manipulating these symbols according to predefined rules.

Main characteristics:

• Explicit knowledge representation: Knowledge is stored as logical statements, rules, semantic networks, or ontologies.
• Rule-based reasoning: The system applies rules such as IF condition THEN action to derive conclusions.
• Interpretability: The reasoning process can usually be traced and explained.
• Dependence on domain experts: Experts must define the knowledge and rules used by the system.
• Limited adaptability: Traditional symbolic systems do not automatically learn effectively from raw data.

Examples:

• Expert systems used for medical diagnosis
• Logic-based theorem provers
• Rule-based chatbots
• Knowledge graphs and planning systems

Symbolic AI performs well in structured domains where rules are clearly known, but it struggles with noisy data, perception, and situations not covered by its rules.

Symbolic AI and Statistical AI differ in how they represent knowledge and solve problems.

Symbolic AI is suitable for structured reasoning, whereas statistical AI is effective for perception and pattern recognition. Modern neuro-symbolic AI attempts to combine the learning ability of statistical methods with the reasoning and interpretability of symbolic methods.

Machine Learning (ML) is a branch of AI that enables computer systems to learn patterns from data and improve their performance without being explicitly programmed for every case.

A typical ML workflow contains the following stages:

1. Problem definition: Identify the task, target output, constraints, and evaluation criteria.
2. Data collection: Gather relevant and representative data from suitable sources.
3. Data preprocessing: Handle missing values, noise, duplicates, inconsistent formats, and outliers.
4. Feature engineering: Select, transform, or construct useful input variables.
5. Data splitting: Divide data into training, validation, and test sets.
6. Model selection: Choose an algorithm appropriate for the task and data.
7. Training: Estimate model parameters by minimizing a loss function.
8. Validation and tuning: Adjust hyperparameters and compare candidate models.
9. Evaluation: Measure performance on unseen test data using suitable metrics.
10. Deployment and monitoring: Integrate the model into an application and monitor accuracy, drift, fairness, and reliability.

The workflow is iterative because evaluation and monitoring may reveal a need for better data, features, or models.

Supervised learning is a machine-learning approach in which a model learns from labeled examples. Each training example contains an input and a known target . The objective is to learn a function such that predicts accurately for unseen inputs.

Major task types:

• Classification: Predicts a discrete category, such as spam or not spam, disease type, or image class.
• Regression: Predicts a continuous value, such as price, temperature, or demand.

Training process:

1. Collect labeled examples .
2. Select a model and initialize its parameters .
3. Generate predictions .
4. Calculate a loss measuring the difference between and .
5. Update to reduce the average loss:

6. Validate the model, tune its hyperparameters, and evaluate it on unseen test data.

Common algorithms include linear regression, logistic regression, decision trees, support vector machines, and neural networks.

Underfitting occurs when a model is too simple to capture the important patterns in the training data. It produces high error on both training and test data and is associated with high bias.

Methods to reduce underfitting:

• Use a more expressive model.
• Add informative features.
• Train for more iterations.
• Reduce excessive regularization.

Overfitting occurs when a model learns the training data too closely, including noise and accidental patterns. It achieves low training error but performs poorly on unseen data and is associated with high variance.

Methods to reduce overfitting:

• Collect more representative training data.
• Apply regularization such as or penalties.
• Use cross-validation for model selection.
• Reduce model complexity.
• Apply pruning to decision trees.
• Use dropout, data augmentation, or early stopping in neural networks.

A good model balances bias and variance and is selected according to its performance on validation data rather than training accuracy alone.

Unsupervised learning is a machine-learning approach that identifies structure or patterns in unlabeled data. The training data contains inputs but no predefined target labels .

Principal tasks:

• Clustering: Divides observations into groups whose members are similar. Examples include customer segmentation and grouping related documents. Algorithms include k-means and hierarchical clustering.
• Dimensionality reduction: Represents high-dimensional data using fewer variables while preserving important information. Examples include visualization and data compression. Principal Component Analysis is a common method.
• Association-rule discovery: Finds relationships among items or events, such as products frequently purchased together.
• Anomaly detection: Identifies unusual observations, such as fraudulent transactions or equipment faults.
• Representation learning: Learns useful features from raw data for later tasks.

Because correct labels are unavailable, evaluation is often more difficult than in supervised learning. Internal measures, domain knowledge, visualization, or downstream-task performance may be used to assess the results.

In supervised learning, the model learns from direct examples of correct answers. In unsupervised learning, it independently discovers patterns. In reinforcement learning, an agent learns through trial and error in an environment. These paradigms may also be combined; for example, unsupervised pretraining can learn representations that are later refined using supervised or reinforcement-learning feedback.

In Reinforcement Learning (RL), an agent repeatedly interacts with an environment and learns behavior that maximizes long-term reward.

Main components:

• Agent: The decision-making entity.
• Environment: The external system with which the agent interacts.
• State : Information describing the current situation.
• Action : A choice available to the agent.
• Reward : Numerical feedback received after an action.
• Policy : A strategy specifying actions in each state.
• Value function: The expected cumulative reward from a state or state-action pair.
• Model: An optional prediction of state transitions and rewards.

The agent seeks to maximize the discounted return:

where is the discount factor.

The exploration-exploitation trade-off concerns whether the agent should try unfamiliar actions to gain information or select the action currently believed to give the highest reward. Excessive exploration reduces immediate reward, while excessive exploitation may prevent discovery of a better strategy. Methods such as -greedy action selection balance the two.

The optimal action-value function represents the maximum expected discounted return obtainable by taking action in state and then following an optimal policy.

The return after taking an action can be separated into the immediate reward and the discounted future return:

After reaching the next state , an optimal agent chooses the action with the greatest action value. Therefore:

If transition probabilities are written explicitly, the equation becomes:

Interpretation:

• is the immediate reward.
• controls the importance of future rewards.
• describes the environment dynamics.
• represents the best future value.

This recursive relationship is the basis of methods such as Q-learning, whose update is:

Here, is the learning rate and the bracketed expression is the temporal-difference error.

Deep Learning is a subfield of machine learning that uses neural networks containing multiple layers to learn hierarchical representations from data. Early layers often learn simple features, while deeper layers combine them into more abstract concepts.

A typical neural-network layer computes:

where contains weights, contains biases, and is an activation function such as ReLU.

Learning process:

1. Forward propagation: The input passes through the layers to produce a prediction.
2. Loss calculation: A loss function measures prediction error.
3. Backpropagation: The chain rule calculates the gradient of the loss with respect to each parameter.
4. Parameter update: An optimizer such as gradient descent updates the parameters:

where is the learning rate.

5. The process is repeated over many batches and epochs.

Deep learning reduces the need for manual feature engineering but often requires substantial data, computation, and careful regularization.

Convolutional Neural Networks (CNNs):

• Apply shared convolutional filters to local regions of data.
• Efficiently capture spatial patterns and translation-related features.
• Commonly used for image classification, object detection, medical imaging, and video analysis.

Recurrent Neural Networks (RNNs):

• Process sequential inputs while maintaining a hidden state.
• Model order and temporal dependencies.
• LSTM and GRU architectures reduce difficulties with long-term dependencies.
• Used for time-series forecasting, speech processing, and sequence modeling.

Transformers:

• Use attention mechanisms to relate different positions in an input sequence.
• Support highly parallel training and capture long-range relationships.
• Form the basis of many language, vision, audio, and multimodal models.
• Used for text generation, translation, summarization, image understanding, and generative AI.

CNNs have a strong locality bias, RNNs process sequences recurrently, and Transformers model relationships through attention. The best architecture depends on the data, computational constraints, and task requirements.

Important deep-learning model types and their applications include:

• Autoencoders: Encode data into a compact representation and reconstruct it. They are used for dimensionality reduction, denoising, anomaly detection, and representation learning.
• Generative Adversarial Networks: Train a generator and discriminator competitively. They are used for image synthesis, style transfer, super-resolution, and data augmentation.
• Diffusion models: Learn to reverse a gradual noising process. They are widely used for generating and editing images, audio, and video.
• Graph Neural Networks: Exchange information between connected nodes. Applications include molecular analysis, recommendation, fraud detection, and social-network analysis.
• Deep reinforcement-learning models: Combine neural networks with reinforcement learning for robotics, games, and control.
• Transformer models: Use attention for language, vision, speech, code generation, and multimodal processing.

Deep learning is also applied to healthcare diagnosis, autonomous driving, industrial inspection, natural-language processing, cybersecurity, and scientific discovery. Model selection should consider the data structure, task, latency, interpretability, and resource requirements.

A foundation model is a large model trained on broad and diverse data that can be adapted to many downstream tasks. Instead of building a separate model from the beginning for every task, developers reuse the capabilities acquired during large-scale pretraining.

Development and use:

• Pretraining: The model learns general representations through objectives such as next-token prediction, masked prediction, or image-text alignment.
• Prompting: Instructions and examples are supplied in the input without changing model parameters.
• Fine-tuning: Parameters are updated using task-specific data.
• Parameter-efficient adaptation: Techniques such as adapters or low-rank adaptation update only a small part of the model.
• Alignment: Human or AI feedback may be used to make outputs more helpful and safer.

Capabilities:

• Transfer learning across many tasks
• Few-shot and zero-shot performance
• Content generation, summarization, reasoning, and retrieval support
• Reuse across language, vision, audio, and code applications

Limitations:

• High training and inference costs
• Bias inherited from training data
• Hallucinated or inaccurate outputs
• Privacy, copyright, security, and misuse concerns
• Limited transparency and difficult evaluation

Foundation models therefore require grounding, monitoring, access controls, and human oversight in high-risk applications.

Transfer learning is the reuse of knowledge learned from one task or dataset to improve performance on another related task. A pretrained model supplies general representations, and a smaller amount of task-specific data is used to adapt it.

Common transfer-learning approaches:

• Use the pretrained model as a fixed feature extractor.
• Fine-tune all model parameters on the target task.
• Fine-tune selected layers while freezing the others.
• Apply parameter-efficient methods such as adapters, prompt tuning, or low-rank adaptation.
• Use in-context learning by providing instructions and examples in the prompt.

Transfer learning is central to foundation models because large-scale pretraining is expensive, while adaptation is comparatively efficient. The pretrained model may already contain useful linguistic, visual, or conceptual patterns. This can:

• Reduce the amount of labeled target data required
• Shorten training time
• Improve performance on specialized tasks
• Allow one base model to support many applications

However, transfer can be harmful when source and target domains differ greatly. Biases, vulnerabilities, and errors from the original model may also transfer to downstream systems.

Multimodal AI refers to systems that process, relate, or generate information across two or more modalities, such as text, images, audio, video, sensor readings, or actions.

Typical processing stages:

1. Modality-specific encoding: Separate encoders convert each input type into numerical representations. For example, a vision encoder processes images and a language encoder processes text.
2. Alignment: Representations of related content are mapped into compatible spaces. An image of a vehicle and its textual description should receive related representations.
3. Fusion: Information is combined using early fusion, late fusion, cross-attention, or a shared transformer.
4. Reasoning or generation: The fused representation supports prediction, retrieval, question answering, or content generation.
5. Decoding: An appropriate decoder produces text, images, audio, or actions.

Applications:

• Visual question answering
• Image captioning and text-to-image generation
• Speech-enabled assistants
• Medical decision support using images and clinical text
• Autonomous systems combining cameras, radar, and other sensors

Major challenges include data alignment, missing modalities, computational cost, conflicting signals, bias, and reliable evaluation across modalities.

Agentic AI refers to AI systems that can pursue goals with a degree of autonomy by observing their environment, planning tasks, taking actions, using tools, and evaluating results. Unlike a system that only produces a single response, an agent may perform a sequence of actions and revise its strategy.

Major components:

• Goal or instruction: Defines the desired outcome and constraints.
• Perception: Collects information from users, documents, sensors, or software environments.
• Reasoning and planning: Breaks a goal into tasks and selects actions.
• Memory: Maintains short-term context and may retrieve long-term knowledge.
• Tool use: Invokes search, databases, APIs, calculators, or other software.
• Action execution: Changes the environment or produces outputs.
• Feedback and reflection: Checks results, detects errors, and updates the plan.
• Control and safety mechanisms: Enforce permissions, budgets, policies, and stopping conditions.

Examples include research assistants, software-development agents, workflow automation systems, and robots. Reliable agents require bounded authority, validated tool inputs and outputs, audit logs, and human approval for consequential actions.

A conventional generative AI assistant generally receives a prompt and generates a response. An agentic system operates in a loop and may plan, call tools, inspect results, maintain memory, and perform multiple actions toward a goal.

Risks of autonomous tool use:

• Executing an incorrect plan due to hallucination
• Prompt injection through untrusted documents or web content
• Unauthorized access or excessive permissions
• Leakage of confidential information
• Repeated actions, uncontrolled cost, or failure to stop
• Cascading failures across connected services

Controls:

• Least-privilege access
• Tool and argument allowlists
• Sandboxed execution
• Human approval for high-impact actions
• Rate, time, and spending limits
• Output validation, monitoring, and audit logs

The degree of oversight should increase with the reversibility and impact of the agent's actions.

Edge AI is the execution of AI workloads on or near the device where data is generated, rather than sending all data to a remote cloud server. Edge devices include smartphones, cameras, vehicles, gateways, industrial controllers, and embedded computers.

Advantages:

• Low latency: Decisions can be made without network round trips.
• Privacy: Sensitive raw data may remain on the local device.
• Offline operation: Applications can continue when connectivity is unavailable.
• Reduced bandwidth: Only selected results need to be transmitted.
• Reliability: Local processing avoids complete dependence on cloud availability.

Limitations:

• Restricted memory, processing power, storage, and energy
• Thermal constraints
• Difficulty updating and monitoring many distributed devices
• Reduced model size or accuracy due to optimization
• Greater physical and software security exposure

Applications:

• Real-time industrial fault detection
• Driver-assistance systems
• Smart surveillance cameras
• Wearable health monitoring
• On-device speech and image recognition
• Agricultural and environmental sensing

Edge AI may be combined with cloud AI so that urgent inference occurs locally while training, aggregation, or complex processing occurs in the cloud.

Tiny AI, often called TinyML, involves running machine-learning models on highly resource-constrained devices such as microcontrollers. Such devices may have very limited memory, low clock speeds, and strict energy budgets.

Difference from general Edge AI:

• Edge AI includes a wide range of local hardware, from powerful edge servers and smartphones to embedded devices.
• Tiny AI focuses on the smallest and most constrained edge platforms.
• Edge devices may run large neural networks with hardware accelerators, while Tiny AI models are usually heavily compressed and optimized.
• Tiny AI often supports always-on sensing using very little power.

Typical techniques:

• Quantization to low-precision integers
• Pruning unimportant weights
• Knowledge distillation
• Compact neural architectures
• Memory-aware inference and optimized kernels

Applications:

• Keyword spotting
• Gesture recognition
• Predictive maintenance using vibration sensors
• Wildlife or environmental monitoring
• Low-power health wearables
• Sensor anomaly detection

Tiny AI enables private, responsive, and low-energy intelligence, but model capacity, memory, battery life, and update mechanisms must be carefully managed.

Deep-learning models are often compressed before deployment on devices with limited computation, memory, and energy.

Major compression techniques:

• Quantization: Replaces high-precision values, such as 32-bit floating-point numbers, with lower-precision representations such as 8-bit integers. It reduces model size and can accelerate inference, but aggressive quantization may reduce accuracy.
• Pruning: Removes weights, channels, or filters that contribute little to predictions. Structured pruning is often easier for hardware to accelerate than irregular pruning.
• Knowledge distillation: Trains a smaller student model to imitate the outputs or internal representations of a larger teacher model.
• Low-rank factorization: Approximates large weight matrices using smaller matrices, reducing parameters and operations.
• Efficient architecture design: Uses compact operations such as depthwise separable convolutions or lightweight attention.
• Operator fusion and compilation: Combines operations and generates hardware-specific code to reduce memory access and runtime overhead.

Trade-offs:

• Greater compression can reduce accuracy and robustness.
• Smaller models may lose rare or complex capabilities.
• Hardware speedups depend on support for the chosen data type and sparsity pattern.
• Optimization may increase development complexity and reduce portability.

A deployment should be evaluated using accuracy, latency, peak memory, energy consumption, model size, and performance on the actual target hardware.