AWS Certified AI Practitioner

C. Fairness

The scenario explicitly mentions concern about 'disparate impact on protected demographic groups' despite high overall performance, which directly relates to the concept of fairness in AI systems.

A. Algorithmic bias

While the root cause might be societal bias reflected in data, the consistent lower scores specifically due to the AI model's processing of zip codes, leading to disparate impact, is an instance of algorithmic bias, where the algorithm itself perpetuates or amplifies existing biases.

C. Prompt engineering combined with Retrieval Augmented Generation (RAG).

Retrieval Augmented Generation (RAG) is a highly effective technique where an LLM's response is generated based on information retrieved from an external knowledge base (like the firm's private legal documents) in real-time, guided by prompt engineering. This avoids costly and time-consuming full model retraining while providing up-to-date and specific information.

A. Interpretability

LIME and SHAP are specific techniques used to explain the predictions of complex AI models, making their decisions understandable. This directly addresses the concept of interpretability, which is about understanding how an AI model makes its decisions.

B. Alignment to specific objectives

The challenge described is about aligning the model's outputs with specific human values, instructions, or desired behaviors (like brand voice, tone, and USPs) that were not perfectly captured in its initial broad pre-training. This 'alignment problem' is a key area in foundation model development, often addressed through techniques like Reinforcement Learning from Human Feedback (RLHF) or specific fine-tuning.

C. Their primary design for unstructured data (text, images) rather than tabular numerical data.

Foundation models, particularly LLMs and Vision Transformers, are predominantly designed and pre-trained on vast amounts of unstructured data like text and images. While adaptations exist, their inherent architecture is less optimized for the direct processing of highly structured, tabular numerical data, which is common in financial transaction analysis, compared to traditional machine learning models tailored for such data.

D. The inference cost and latency associated with each model.

Given a tight budget and limited GPU resources, the inference cost (which directly relates to the computational resources needed to run the model) and latency (how quickly it responds) are critical. Larger models (like Model A) typically have higher inference costs and latency, while smaller or fine-tuned models (like Model B) might be more efficient and cost-effective for deployment, especially with resource constraints.

A. It can process and integrate information from different data types (e.g., genomic sequences, protein structures, research papers) simultaneously.

Multi-modal foundation models are designed to process and understand information across multiple data modalities (e.g., text, images, audio, structured data) simultaneously. For genomic research, this allows the model to integrate insights from genomic sequences, protein structure images, and relevant scientific literature, leading to a more comprehensive analysis and potentially novel discoveries.

C. Human-in-the-Loop (HITL)

Establishing a process for human oversight, review points, and override mechanisms directly implements the Human-in-the-Loop (HITL) concept, where human judgment is integrated into an AI system's workflow.

B. The capabilities of very large models are not always predictable from smaller versions.

Emergent abilities are new capabilities that appear in larger foundation models but are not present in smaller models and were not explicitly trained for. This implies that scaling up models can lead to unpredictable, non-linear increases in performance and functionality, making it difficult to fully foresee their potential from smaller-scale experiments.

D. Privacy

Anonymization, pseudonymization, and secure data storage are all practices aimed at protecting sensitive personal data, which directly falls under the umbrella of AI privacy.

B. Perform domain-adaptive pre-training on the company's proprietary documentation.

Domain-adaptive pre-training (DAPT), also known as continued pre-training, involves taking a pre-trained foundation model and further pre-training it on a large corpus of domain-specific unlabeled data. This allows the model to learn the specific vocabulary, jargon, and nuances of the new domain without losing its general capabilities, making it highly effective for specialized tasks like understanding proprietary software documentation.

D. Robustness

The effort to prevent the AI system from being 'easily manipulated by malicious actors' and to handle 'slight modifications to harmful content' directly addresses the concept of robustness, ensuring the model maintains performance under adversarial conditions.

D. Their training objective to predict the next token, not to verify factual accuracy.

Foundation models, especially Large Language Models, are trained with an objective to predict the next most probable token based on their vast training data. This probabilistic generation does not inherently include a mechanism for factual verification, leading to instances where the model generates coherent but incorrect information (hallucinations).

B. Their pre-training on broad data, enabling strong generalization and few-shot learning.

Foundation models are pre-trained on vast and diverse datasets, which allows them to learn general representations that can be fine-tuned for specific downstream tasks with minimal labeled data (few-shot learning) or even zero-shot learning, making them highly effective for content moderation where specific examples might be scarce.

A. Deploying the foundation model on-premises or in a private cloud environment.

Deploying the foundation model on-premises or within a private cloud environment allows the healthcare provider to maintain complete control over their sensitive patient data, ensuring it never leaves their secure, compliant infrastructure. This addresses strict privacy regulations like HIPAA more effectively than public cloud APIs or anonymized data, which might still carry residual risks.

D. Large Language Model (LLM)

Large Language Models (LLMs) are specifically designed and trained on vast amounts of text data to understand, generate, and process human language, making them ideal for chatbot applications requiring coherent text responses.

A. Recall (Sensitivity)

In a scenario with a rare disease where failing to detect positive cases (false negatives) has severe consequences, Recall (Sensitivity) is the most critical metric. It measures the proportion of actual positive cases that were correctly identified, ensuring that few sick individuals are missed.

B. Implementing a Retrieval Augmented Generation (RAG) system.

Retrieval Augmented Generation (RAG) is designed to address this exact problem. It allows the foundation model to retrieve up-to-date and specific information from an external, dynamic knowledge base (like company policies) and then use that retrieved context to generate a more accurate and relevant answer. This avoids costly full retraining and ensures the model's responses are grounded in current data.

C. Transparency

The company's desire to understand 'why' the AI made specific recommendations directly relates to transparency, which involves making AI system operations and decisions understandable.

B. Collect additional, representative data for the underrepresented group and retrain the model.

The most effective way to address poor performance due to underrepresentation is to gather more representative data for that specific group and retrain the model, directly improving its ability to generalize to that population.

A. Ethical Impact Assessment (EIA)

A proactive evaluation process specifically designed to identify 'potential negative societal impacts, ethical dilemmas, and risks of discrimination' before deployment is known as an Ethical Impact Assessment (EIA) or AI Impact Assessment.

D. Accountability

Defining 'who is responsible for model errors, data breaches, or unintended negative consequences' directly aligns with the concept of accountability in Responsible AI, which establishes clear ownership and responsibility for AI system outcomes.

C. Algorithmic bias

Algorithmic bias occurs when a model's outputs are systematically prejudiced or unfair due to biases present in its training data. In this case, the model's preference for bubble sort, even when less optimal, stems from an imbalanced representation of sorting algorithms in its pre-training data, reflecting a bias towards simpler or more common examples found online.

D. Generalization capabilities

Foundation models possess strong generalization capabilities, meaning they can perform well on tasks or data distributions that differ from their original training data, often with little to no fine-tuning (zero-shot or few-shot learning). This characteristic allows the content moderation system to adapt to new types of inappropriate content quickly without extensive retraining, as the model can generalize from its vast pre-training knowledge.