ISTQB - AI Testing(CT-AI) - 1. Introduction to AI -Test 03

How has the definition of Artificial Intelligence (AI) evolved since the 1950s?. It has become narrower, focusing only on machine learning. It has remained unchanged, centered on imitating human beings. It has evolved significantly; a modern definition is "The capability of an engineered system to acquire, process and apply knowledge and skills.". It now exclusively refers to superintelligent systems.

What is an example of the "AI effect" regarding expert systems from the 1970s and 1980s?. They are now considered more intelligent than modern AI. They were considered AI then, but many do not consider them as such now, showcasing the shifting perception of AI. Their definition as AI has been reinforced over time. The AI effect made them obsolete very quickly.

Which type of AI are voice assistants an example of?. General AI. Super AI. Narrow AI. Theoretical AI.

What is the primary capability of a General AI system?. To perform one specific task with superhuman accuracy. To have general (wide-ranging) cognitive abilities similar to humans, enabling reasoning and understanding of the environment. To access and process all human knowledge from the web. To operate without any prior programming or data.

The transition point from General AI to Super AI is known as: The AI Effect. Machine Learning. Technological Singularity. Deep Learning.

How is an AI-based image processor trained to identify cats, according to the syllabus?. By programmers writing explicit rules for "cat-like" features. It is trained with a set of images known to contain cats, and the AI determines on its own what patterns to use. By using fuzzy logic and rule engines exclusively. It comes pre-programmed with knowledge of all animal species.

Which statement accurately compares AI-based systems using ML and conventional systems?. Conventional systems are generally less understandable by humans than AI-based systems. AI-based systems using ML rely on explicitly programmed if-then-else logic for their learning. Conventional systems use imperative languages, while AI systems using ML use patterns in data to determine future reactions. Only conventional systems can be implemented with technologies like decision trees.

Which of the following is listed in the syllabus as an AI technology?. Cascading Style Sheets (CSS). Random forest. Virtual Private Networks (VPN). Microservices architecture.

What is the Microsoft Cognitive Toolkit (CNTK)?. A high-level open-source API written in Python. An open-source deep-learning toolkit. An ML framework based on data flow graphs from Google. A suite of tools by IBM for AI solutions.

AI development frameworks support a range of activities. Which activity is NOT explicitly listed as supported by them in section 1.5?. Data preparation. Algorithm selection. Ethical review and compliance checking. Compilation of models to run on various processors.

Which hardware feature refers to the ability to perform calculations concurrently on many processing units?. Low-precision arithmetic. Large data structure support. Massively parallel (concurrent) processing. In-memory processing.

Why might a model performing speech recognition run on a low-end smartphone, yet require cloud computing for training?. Smartphones have superior processing power for training. Training ML models is often computationally intensive, requiring significant power, while running a trained model (inference) can be less demanding. Cloud computing is only used for storing trained models. Low-end smartphones cannot connect to the internet for model updates.

What kind of hardware are Google TPUs (Cloud Tensor Processing Units)?. Neuromorphic processors. Application-specific integrated circuits (ASICs) that can be accessed by users on the Google Cloud. Standard GPUs rebranded by Google. CPUs with exceptionally high clock speeds.

Which company provides Nervana neural network processors for deep learning?. Apple. Huawei. NVIDIA. Intel.

Which of these is an example of AIaaS provided by Google Cloud?. An AI chatbot priced per monthly active user. Document-based AI including a form parser and OCR, priced per page. A review of ML Java code priced per line of code. AI cloud search priced by search units.

What is a common reason for an AIaaS SLA to RARELY define ML functional performance metrics like accuracy?. Accuracy is not important for AIaaS. Defining and guaranteeing specific ML performance metrics like accuracy in a general SLA is complex and depends heavily on the input data and specific use case. SLAs tend to cover more controllable aspects like uptime. These metrics are always covered in a separate, mandatory contract. Service providers prefer to offer credits rather than guarantee performance.

Why might an AI-based system depending on AIaaS be limited to relatively low-risk applications?. AIaaS is generally more expensive than in-house development for high-risk applications. Most AIaaS contracts provide limited liability (other than fees paid), making them unsuitable where loss of service would be too damaging. AIaaS providers typically lack the expertise for high-risk scenarios. High-risk applications always require on-premise hardware.

What is one of the main reasons organizations opt to use pre-trained models?. To ensure the model is always up-to-date with the very latest research. It can be expensive to train ML models, requiring significant human and computing resources that many organizations lack; pre-trained models offer a cheaper, often more effective alternative. Pre-trained models are guaranteed to be perfectly documented. They are only available for highly specialized, niche technologies.

For which technologies are pre-trained models mentioned as being available (though limited)?. Fuzzy logic and rule engines. Neural networks and random forests. Deductive classifiers and case-based reasoning. Linear regression and logistic regression.

If an image classifier is needed, what is an alternative to training it on a large dataset like ImageNet from scratch?. Manually programming the classification rules. Using an existing model that has already been trained on this dataset (a pre-trained model). Developing a new, smaller dataset specific to the task. Using a conventional programming approach without ML.

In the context of transfer learning with an image classifier, if the early layers identify basic features like lines, what might later layers do?. Perform data pre-processing. Perform more specialized tasks, like differentiating between building architectural types. Optimize the model's memory usage. Convert the image into a textual description.

What is an example of a pre-trained NLP model mentioned in the syllabus?. ImageNet. AlexNet. Google's BERT. MobileNet.

What risk arises from differences in data preparation steps between a pre-trained model's original development and its new use?. It may improve the model's transparency. The resulting functional performance may be impacted. It makes the model less sensitive to adversarial attacks. It ensures better documentation of the model.

How can thorough documentation for a pre-trained model help mitigate risks?. It guarantees the model is free from all defects. It can help mitigate risks such as lack of transparency, undocumented shortcomings, or inherited biases. It automatically updates the model to fix vulnerabilities. It ensures the model will perform perfectly for any new task.

Which standard is specifically mentioned as relevant for AI in safety-related automotive systems?. GDPR. ISO/IEC JTC 1/SC42 guidelines. ISO 26262. The AI Quality Metamodel by DIN.

According to the syllabus, the use of standards is normally made mandatory by: The AI Effect. The developers of AI frameworks. Legislation or contract. Academic research institutions.

Which of the following is NOT a popular AI development framework listed in section 1.5?. CNTK. Keras. Altera. Scikit-learn.

What is a key feature of Al-specific hardware solutions like ASICs and SoCs, making them suitable for edge computing?. They primarily rely on high-precision arithmetic. They typically have few cores and focus on general-purpose operations. Features like multiple cores, special data management, and in-memory processing, while training is done in the cloud. Their primary advantage is extremely high clock speeds.

IBM Watson Studio is described as: An open-source deep-learning toolkit. A high-level open-source API for Python. A suite of tools that support the development of AI solutions. An open-source ML library operated by Facebook.

What does the IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems work on?. Developing hardware like GPUs and TPUs. Publishing ML libraries for Python. A range of standards on ethics and AI. Mandating the use of specific AI development frameworks.