Topic A4.1 Machine Learning Fundamentals & Hardware Deployment

Flashcards covering IB Computer Science SL Topic A4.1 on Machine Learning Fundamentals, Learning Paradigms, Application Scenarios, and Hardware Deployment.

In traditional programming, what is the formula that represents how inputs and explicit rules produce an output?

$$\text{Data} + \text{Rules} \longrightarrow \text{Output}$$

In machine learning, what is the formula that represents how inputs and expected outputs generate rules?

$$\text{Data} + \text{Outputs} \longrightarrow \text{Rules (Model)}$$

What is the core characteristic of supervised learning?

It is trained on labeled data, where each input instance $X$ is paired with a verified target output $Y$ (ground truth).

What are the two subcategories of supervised learning mentioned in the text, along with their output types?

Classification (output is a discrete categorical variable) and Regression (output is a continuous numerical quantity).

What is the core characteristic of unsupervised learning?

It is trained on unlabeled data ($X$ only) with no pre-assigned ground-truth target outputs.

What are the three subcategories of unsupervised learning listed in the text?

1. Clustering, 2. Association Rule Learning, 3. Anomaly Detection.

In Association Rule Learning, what co-occurrence percentage example is given in the text?

"If Item A is bought, Item B is also bought $80\%$ of the time".

What are the four components of the Reinforcement Learning (RL) process described in the text?

State ($S$), Action ($A$), Reward/Penalty ($R$), and Policy Update.

Which two algorithms does Deep Learning use to tune millions or billions of network weights?

Backpropagation and Gradient Descent algorithms.

What core process is involved in Transfer Learning when adapting a pre-trained network to a new target task?

The early layers of the pre-trained model are frozen, and the final classification layers are replaced and fine-tuned on a small dataset specific to the target domain.

Why is Deep Learning / Transfer Learning used for Medical Imaging Diagnostics such as detecting tumors in MRI scans?

CNNs (Convolutional Neural Networks) excel at spatial feature extraction from complex pixel matrices, and transfer learning allows high accuracy even when specialized clinical training scans are scarce.

What primary ML approach is used for Sentiment Analysis, and into what categories are text sequence features mapped?

Supervised Learning (NLP); mapped directly to discrete categorical sentiment labels (Positive, Negative, Neutral) using annotated corpora.

Why is Unsupervised Learning (Clustering) used for Customer Segmentation?

Marketing teams do not have pre-existing segment labels, so the algorithm groups users based on purchasing behavior similarities.

Why is Anomaly Detection applied in Credit Card Fraud Detection?

Fraud patterns change constantly, and anomaly detection flags transactions that diverge drastically from a user's baseline spending profile.

In the Problem-Solving & Paradigm Selection Framework, how is the choice made during Step 1 (Check Data Labeling)?

If labeled ground-truth data exists, proceed to Supervised Learning or Transfer Learning. If NO labeled data exists, proceed to Unsupervised Learning or Reinforcement Learning.

What is the computational profile and hardware focus of the Training Phase in ML deployment?

Computational profile: Highly iterative matrix math, heavy backpropagation calculations, and massive parallel throughput requirements. Hardware focus: Multi-GPU clusters, Cloud Tensor Processing Units (TPUs), maximum memory bandwidth, and high VRAM capacity.

What is the definition and hardware focus of the Inference Phase in ML deployment?

Definition: Running a fully trained, static model on new real-world input to generate instant predictions. Hardware focus: Edge devices, Neural Processing Units (NPUs), low-power microcontrollers, and optimized execution speed per watt.

How do CPUs and GPUs differ in core architecture according to the hardware comparison table?

CPUs have few powerful cores optimized for sequential thread execution ($4$–$64$ cores), while GPUs have thousands of smaller, simpler cores optimized for simultaneous parallel math ($1000$+ cores).

What processing tasks is a Graphics Processing Unit (GPU) optimized for in contrast to a CPU?

GPUs are optimized for high throughput, parallel matrix multiplication, and vector arithmetic, whereas CPUs are optimized for complex logic, conditional branching (if/else), and low sequential latency.

What are Tensor Processing Units (TPUs) and what is their primary use?

Custom Application-Specific Integrated Circuits designed specifically for matrix math operations, primarily used for large-scale cloud deep learning training (e.g., Large Language Models).

What are the characteristics of ASICs (Application-Specific Integrated Circuits)?

They are microchips permanently hardwired for one single computational workload. They are non-reconfigurable, offering maximum processing efficiency and lowest power draw for fixed, high-volume tasks.

What are Field-Programmable Gate Arrays (FPGAs) and why are they used?

They are reconfigurable microchips that can be rewritten at the hardware level after manufacturing, making them ideal for low-latency custom inference where algorithms update frequently.

What primary hardware choice and justification factor are specified for an Autonomous Vehicle Driving System?

Hardware Choice: Embedded High-Performance Edge System (e.g., dual NPU/GPU modules). Justification: Latency & Safety, requiring instant millisecond response where local execution avoids network failure risks.

What primary hardware choice and justification factor are given for a Smart Health Tracker Wearable?

Hardware Choice: Ultra-low power Microcontroller with NPU (TinyML). Justification: Power & Size, as it must run on a coin-cell battery with strict thermal and memory footprint constraints.

What hardware choice is recommended for a Custom High-Frequency Financial Trading Model and why?

Hardware Choice: FPGA. Justification: Custom Latency, as reconfigurable hardware logic bypasses standard OS overhead for sub-microsecond execution.