AIP-210 Sample Questions Answers

Emergent bias is a type of bias that arises when an AI model encounters new or different data or scenarios that were not present or accounted for during its training or development. Emergent bias can cause the model to make inaccurate or unfair predictions or decisions, as it may not be able to generalize well to new situations or adapt to changing conditions. One possible cause of emergent bias is seasonality, which means that some variables or patterns in the data may vary depending on the time of year. For example, if an AI model for merchandise sales prediction was trained in winter and applied in summer, it may produce biased results due to differences in customer behavior, demand, or preferences.

IBM Watson is an AI technology that can take a question in natural language and return a precise answer to the question. IBM Watson is a cognitive computing system that can understand natural language, generate hypotheses, and provide evidence-based answers. IBM Watson can be applied to various domains and industries, such as healthcare, education, finance, or law.

Scalability and explainability are two criteria that are essential for ML models to achieve before deployment. Scalability is the ability of an ML model to handle increasing amounts of data or requests without compromising its performance or quality. Scalability can help ensure that the model can meet the demand and expectations of users or customers, as well as adapt to changing conditions or environments. Explainability is the ability of an ML model to provide clear and intuitive explanations for its predictions or decisions. Explainability can help increase trust and confidence among users or stakeholders, as well as enable accountability and responsibility for the model’s actions and outcomes.

Generative adversarial networks (GAN) are a type of AI technology that can create fake videos, images, audio, or text that are realistic and indistinguishable from real ones. GAN consist of two neural networks: a generator and a discriminator. The generator tries to produce fake samples from random noise, while the discriminator tries to distinguish between real and fake samples. The two networks compete against each other in a game-like scenario, where the generator tries to fool the discriminator and the discriminator tries to catch the generator. Through this process, both networks improve their abilities until they reach an equilibrium where the generator can produce convincing fakes.

Word embedding is a task in natural language processing (NLP) where words are converted into numerical vectors that represent their meaning, usage, or context. Word embedding can help reduce the dimensionality and sparsity of text data, as well as enable various operations and comparisons among words based on their vector representations. Some of the common methods for word embedding are:

• One-hot encoding: One-hot encoding is a method that assigns a unique binary vector to each word in a vocabulary. The vector has only one element with a value of 1 (the hot bit) and the rest with a value of 0. One-hot encoding can create distinct and orthogonal vectors for each word, but it does not capture any semantic or syntactic information about words.
• Word2vec: Word2vec is a method that learns a dense and continuous vector representation for each word based on its context in a large corpus of text. Word2vec can capture the semantic and syntactic similarity and relationships among words, such as synonyms, antonyms, analogies, or associations.
• GloVe: GloVe (Global Vectors for Word Representation) is a method that combines the advantages of count-based methods (such as TF-IDF) and predictive methods (such as Word2vec) to create word vectors. GloVe can leverage both global and local information from a large corpus of text to capture the co-occurrence patterns and probabilities of words.

Lasso regression and ridge regression are both types of linear regression models that can handle high-dimensional and categorical data. They use regularization techniques to reduce the complexity of the model and avoid overfitting. Lasso regression uses L1 regularization, which adds a penalty term proportional to the absolute value of the coefficients to the loss function. This can shrink some coefficients to zero and perform feature selection. Ridge regression uses L2 regularization, which adds a penalty term proportional to the square of the coefficients to the loss function. This can shrink all coefficients towards zero and reduce multicollinearity. References: [Lasso (statistics) - Wikipedia], [Ridge regression - Wikipedia]

Precision and recall are two measures that can evaluate the performance of a classifier, especially when the data is imbalanced. Precision is the ratio of true positives (correctly predicted positive cases) to all predicted positive cases. Recall is the ratio of true positives to all actual positive cases. Precision and recall can help assess how well the classifier can identify the positive cases (the disease) and avoid false negatives (missed diagnosis) or false positives (unnecessary treatment).

The transform-design pattern for ML pipelines aims to separate inputs from features, encapsulate the processing steps of ML pipelines, and represent steps in the pipeline with a DAG. These goals help to make the pipeline modular, reusable, and easy to understand. The transform-design pattern does not seek to isolate individual steps of ML pipelines, as this would create entanglement and dependency issues. It also does not transform the output data after production, as this would violate the principle of separation of concerns. 

The General Data Protection Regulation (GDPR) is a privacy-focused law that an AI practitioner should adhere to while designing and adapting an AI system that utilizes personal data. The GDPR applies to any organization that processes personal data of individuals in the European Union (EU), regardless of where the organization is located. The GDPR grants individuals rights over their personal data, such as the right to access, rectify, erase, restrict, or object to its processing. The GDPR also imposes obligations on organizations that process personal data, such as the duty to obtain consent, conduct data protection impact assessments, implement data protection by design and by default, and ensure accountability and transparency. The GDPR also addresses some specific issues related to AI, such as automated decision-making, profiling, and data portability.

Latency is the time delay between a request and a response. Latency can affect the performance and user experience of an application, especially when real-time or near-real-time responses are required. Deploying a deep learning model as an embedded model on edge devices can reduce latency, as the model can run locally on the device without relying on network connectivity or cloud servers. Edge devices are devices that are located at the edge of a network, such as smartphones, tablets, laptops, sensors, cameras, or drones.

The optimal value of K is the one that minimizes the inertia or within-cluster sum of squares, while avoiding too many clusters that may overfit the data. The elbow plot shows a sharp decrease in inertia from K = 1 to K = 2, and then a more gradual decrease from K = 2 to K = 3. After K = 3, the inertia does not change much as K increases. Therefore, the elbow point is at K = 3, which is the optimal value of K for this data. References: How to Run K-Means Clustering in Python, K-means clustering - Wikipedia

Ridge regression is a type of linear regression that adds a regularization term to the loss function to reduce overfitting and improve generalization. Ridge regression is fast and requires low memory and low processing power, as it only involves solving a system of linear equations. Ridge regression can also handle multicollinearity (high correlation among predictors) by shrinking the coefficients of correlated predictors.

Ridge regression is a type of regularization technique that can help reduce collinearity among independent variables. It does this by adding a penalty term to the ordinary least squares (OLS) objective function, which shrinks the coefficients of highly correlated variables towards zero. This reduces the variance of the coefficient estimates and improves the stability and accuracy of the regression model. References: Multicollinearity in Regression Analysis: Problems, Detection, and Solutions - Statistics By Jim, A Beginner’s Guide to Collinearity: What it is and How it affects our regression model - StrataScratch

A false negative is an error where a positive case (belonging to the target class) is incorrectly predicted as negative (not belonging to the target class). A false negative rate is the ratio of false negatives to all actual positive cases. A low false negative rate means that most of the positive cases are correctly identified by the classifier.

For a diagnostic tool that can diagnose a particular disease so that individuals with the disease can receive treatment, it is of primary importance that it has a low false negative rate. This is because false negatives can have serious consequences for patients who have the disease but do not receive treatment, such as increased risk of mortality or complications. A low false negative rate can ensure that most patients who have the disease are diagnosed correctly and receive timely treatment.

Text vectorization is a technique that converts text into numerical vectors that can be used by machine learning models. Text vectorization can use different methods to represent text features, such as word frequency, word order, word meaning, or word context. Some of the common text vectorization methods are:

• TF-IDF: TF-IDF (term frequency-inverse document frequency) is a method that assigns a weight to each word based on its frequency in a document and its rarity across a collection of documents. TF-IDF can capture the importance and relevance of words for a given topic or domain, but it does not consider the order or meaning of words.
• Word2vec: Word2vec is a method that learns a vector representation for each word based on its context in a large corpus of text. Word2vec can capture the semantic and syntactic similarity and relationships among words, as well as preserve the order of words.

For an English-to-Spanish translation machine, using Word2vec would be appropriate and correctly defined, because in translation machines, we need to consider the order of the words, as well as their meaning and context.

A type 1 error in statistical hypothesis testing is when the null hypothesis is true, but is rejected. This means that the test falsely concludes that there is a significant difference or effect when there is none. The probability of making a type 1 error is denoted by alpha, which is also known as the significance level of the test. A type 1 error can be reduced by choosing a smaller alpha value, but this may increase the chance of making a type 2 error, which is when the null hypothesis is false but fails to be rejected. References: [Type I and type II errors - Wikipedia], [Type I Error and Type II Error - Statistics How To]

The Adversarial ML Threat Matrix is an open framework designed to help detect, respond to, and remediate threats in ML systems. The Adversarial ML Threat Matrix is inspired by the MITRE ATT&CK® Matrix1, which is a framework for describing cyberattacks across various stages of an attack lifecycle. The Adversarial ML Threat Matrix adapts this framework to address specific threats and vulnerabilities in ML systems, such as data poisoning, model stealing, model evasion, or model inversion2. The Adversarial ML Threat Matrix provides a structured way to organize and classify adversarial techniques, tactics, procedures, examples, and mitigations for ML systems2.

Neural network models are suitable for classification problems with a large number of features, because they can learn complex and non-linear patterns from high-dimensional data. They can also handle image data, which is likely to be the input for the human face detection problem. Neural networks can also be trained using transfer learning, which can leverage pre-trained models on similar tasks and improve the accuracy and efficiency of the model. References: [Neural network - Wikipedia], [Transfer Learning - Machine Learning’s Next Frontier]

Security measures can be applied in different ML workflow stages to defend them against malicious activities, such as data theft, model tampering, or adversarial attacks. Some of the security measures are:

• Launch ML Instances In a virtual private cloud (VPC): A VPC is a logically isolated section of a cloud provider’s network that allows users to launch and control their own resources. By launching ML instances in a VPC, users can enhance the security and privacy of their data and models, as well as restrict the access and traffic to and from the instances.
• Use data encryption: Data encryption is the process of transforming data into an unreadable format using a secret key or algorithm. Data encryption can protect the confidentiality, integrity, and availability of data at rest (stored in databases or files) or in transit (transferred over networks). Data encryption can prevent unauthorized access, modification, or leakage of sensitive data.
• Use Secrets Manager to protect credentials: Secrets Manager is a service that helps users securely store, manage, and retrieve secrets, such as passwords, API keys, tokens, or certificates. Secrets Manager can help users protect their credentials from unauthorized access or exposure, as well as rotate them automatically to comply with security policies.

Anomaly detection is an unsupervised learning technique that identifies outliers or abnormal patterns in data, which can be useful for fraud detection. Anomaly detection algorithms can learn the normal behavior of transactions and flag the ones that deviate significantly from the norm, indicating possible fraud.

Performance testing is a type of testing that should be performed at the production level before deploying a newly retrained model. Performance testing measures how well the model meets the non-functional requirements, such as speed, scalability, reliability, availability, and resource consumption. Performance testing can help identify any bottlenecks or issues that may affect the user experience or satisfaction with the model. References: [Performance Testing Tutorial: What is, Types, Metrics & Example], [Performance Testing for Machine Learning Systems | by David Talby | Towards Data Science]

Semi-supervised learning is an approach that is best if a limited portion of your training data is labeled. Semi-supervised learning is a type of machine learning that uses both labeled and unlabeled data to train a model. Semi-supervised learning can leverage the large amount of unlabeled data that is easier and cheaper to obtain and use it to improve the model’s performance. Semi-supervised learning can use various techniques, such as self-training, co-training, or generative models, to incorporate unlabeled data into the learning process.

Caffe is a deep learning framework that is designed for speed and modularity. It can process large datasets of images efficiently and supports various types of neural networks. It also has a large and active developer community that contributes to its code base and documentation. Caffe is suitable for image processing tasks such as classification, segmentation, detection, and recognition

The Beta value in an A/B test is the probability of making a type II error, which is failing to reject the null hypothesis when it is false. The statistical power of a test is the probability of correctly rejecting the null hypothesis when it is false, which is equal to 1 - Beta. References: Formulas for Bayesian A/B Testing – Evan Miller, The Practical Guide To AB testing statistics | Convertize