hs1501 quiz .pdf

S2: Capabilities: Language 1. Which of the following best describes the current level of AI-powered text chatbots? They are always honest with what they do not have information about. They are perfect in detecting and handling the user’s sentiment. They sometimes produce text with obscure...

S2: Capabilities: Language 1. Which of the following best describes the current level of AI-powered text chatbots? They are always honest with what they do not have information about. They are perfect in detecting and handling the user’s sentiment. They sometimes produce text with obscure grammatical errors. They chat fluently, but the facts they claim are not always accurate. This is referring to the section about chatbots in §2.10 and the part about hallucination in §2.11. Sentiment analysis can sometimes be wrong, as seen in §2.2. As the exercises in §2.9 shows, AI-powered text chatbot almost never make grammatical errors, let alone obscure ones. 2. Which of the following AI capabilities would be best for extracting key theme topics from a business article? Language translation. Sentiment analysis. Named entity recognition. Text categorization. 3. Which of the following is the most significant difference between conventional Google-style Internet search and current AI-assisted Internet search like perplexity.ai? Conventional Google-style search relies mostly on approximate keyword matching, while AI-assisted search takes into account the conceptual meanings of words and the intent behind the query. Conventional Google-style search ranks search results based on the number of high-quality links to them, while AI-assisted search ranks search results based on the perceived reliability of the websites. Conventional Google-style search is known to return more objective results because it uses more deterministic algorithms, while AI-assisted search may show bias that is inherent in the training data. Conventional Google-style search works well for the layman who needs to get only a general idea, while AI-assisted search is more suited for the professional who requires also the nitty-gritty details. Some of the points we expect students to find out in the exercise on perplexity.ai in §2.4 is that there is less keyword matching, and no ranking of search results. Both Google-style search and AI-assisted search reflect what is on the Internet, whether it is biased or detailed. 4. Which of the following is the LEAST fruitful way in which you can use chatbots like ChatGPT for your HS1501 Mini-Project 1? Refining the tone and the style of the writing. Exploring new ideas and alternative viewpoints. Explaining complex concepts in simple terms. Fact-checking claims found on the Internet. ChatGPT is not good for fact-checking because it hallucinates; see the relevant point in §2.11. Some potential uses of such chatbots are mentioned in §2.10. 5. When would sentiment analysis be challenging at present? When the text involves colloquial expressions and proverbs. When the text involves sarcasm and language nuances. When the text consists of multiple sentences. When the text is too short, usually meaning less than 100 words. This is referring to the bullet point about sentiment analysis in §2.11. One can see from the Hugging Face exercise in §2.2 that AI-powered sentiment analysis can now handle colloquial expressions, proverbs, short text, and multiple sentences. Logically, if multiple sentences per se were an issue for a particular AI, then replacing all full stops by semicolons (or an appropriate punctuation in the language) would easily solve it.  6. How could Al help the music industry directly? Replace musicians with robot singers and dancers. Generate new lyrics and compose new songs. Personalise YouTube recommendations and create personalised playlists. Reduce the bandwidth of Netflix movie-streaming services. 7. What do chatbots require Al to do? Al enables chatbot users to chat with other people on social media platforms. Al makes chatbots multilingual. Al helps chatbots go beyond a predetermined set of answers. Al allows chatbots to produce (not only text but) images as responses. 8. When would sentiment analysis be challenging at present? When the text is too short, usually meaning less than 100 words. When the text consists of multiple sentences. When the text involves sarcasm and language nuances. When the text involves colloquial expressions and proverbs. 9. Why is it harder to make an LLM that translates Icelandic compared to Spanish or Chinese? Icelandic words are typically very long. There are limited digital resources available in Icelandic. The grammatical structure of Icelandic is complex. Many syllables in Icelandic are silent. LLM MOD A 1. What is generative AI? A type of artificial intelligence that can only recognize speech and convert it into text. A type of artificial intelligence that can only classify or categorize database tabular data. A type of artificial intelligence that can only analyze data and make predictions about future events. A type of artificial intelligence that can create new content, such as images, music, or text, by learning patterns from existing data. 2. Why was ChatGPT so successful and became so quickly popular? Because it is the largest language model ever created. Because it is the first language model to use the Transformer architecture. Because it is the only language model that is capable of generating creative and imaginative responses. Because it is a highly capable and versatile language model that can be easily used by anyone to generate coherent responses to a wide range of questions S3: Capabilities: Vision 1. Which are some shortcomings of current facial recognition AI? The person must remain stationary for an extended period of time for the face to be recognized accurately. It takes a long time and a lot of computational power to run facial recognition AI even when executed in the cloud. Facial recognition AI currently requires expensive 3D cameras and infrared cameras to give accurate results. Dark skin colour, femaleness, hair style, and make-up may negatively affect the accuracy of facial recognition. This is referring to Prof. Yu’s video in §3.3. Ordinary cameras are already enough for facial recognition, and the person does not need to remain stationary for very long. It may take a lot of time and power to train the facial recognition AI, but not to run it, especially with the help of cloud infrastructures. 2. How are AI-based text recognition algorithms superior to non-AI-based ones? It can produce a text description of an image even when the image contains no text. It can read text in any colour. It can accurately figure out text in photos even when the view is blocked. It can read highly curved text in a variety of fonts. This was demonstrated in §3.1. Non-AI-based text recognition algorithms are also able to read text in any colour (possibly after a simple colour change that does not involve AI). When the view is blocked, it is not possible to tell accurately what the text is from a photo. Visual question answering techniques can produce a text description of an image, text recognition cannot.  3. Which of the following is the least challenging for current text-to-video AI to produce? Physically realistic videos. Videos in high resolution. Consistent characters. Videos that are long in duration. These are some observations from watching the video in §3.6. It is of rather high resolution. However, it is mostly made up of small clips, suggesting that they are still struggling to produce long videos. The choice of a balloon head subtly avoids the problem of having to generate a consistent character (and also the problem of generating a character who speaks), but the balloon is not moving in a physically realistic way. 4. Which of the following applications of AI’s vision capabilities requires the processing time to be in hundredths of a second? Identifying cancer cells. Colourizing old movies. Immigration clearance. Autonomous driving. This is mentioned in §3.2. 5. Which of the following applications relies chiefly on AI’s capability to recognize and count objects visually? Monitoring the number of people in the queue. Detecting stolen objects at the supermarket exit. Keeping track of who ordered what at a food stall in a hawker centre. Counting the number of animals in the zoo that are sick. Counting the number of people in the queue is a typical application of object recognition. One cannot distinguish between stolen objects and bought objects from their appearances. Similarly, one often cannot tell from the appearance whether an animal is sick. Object recognition/counting AI usually cannot tell between different people, and thus cannot keep track of orders. 6. Which of the following is an application of Al image/video generation in education? Learning progress tracked and displayed using data analytics. Students' questions answered by a chatbot round the clock. Lecture content delivered by a digital instructor in a video. Tuition fee collected according to attendance record. 7. Which of the following tasks needs semantic segmentation technology? Recognizing street signs and names of shops in the street. Counting airplanes in an airport using satellite imagery. Identifying car lanes and neighbouring cars for an autonomous vehicle. Locating timepoints at which a movie changes scene. 8. In which of the following Al applications would reverse image search be required? A shopper sees how a piece of clothing looks on the shopper by taking a photo of the item. A shopper gets a smart factory to reproduce l...] a pair of shoes [...] taking a photo of the item. A shopper checks the expiry date of a food item in a physical store by taking a photo of the item. A shopper compares prices on different e-commerce sites [.] physical store by taking a photo [...] S4: Capabilities: Robots 1. Which is an application of AI-powered RPA? Predicting which rubbish bins fill up more quickly at what time in the shopping mall, and clearing those that are predicted to be full timely. Personalizing the taste of bubble tea according to the customer’s personality, and delivering the personalized product to the customer. Sending out e-mails to each student in the course individual venue, seating, and timing information for the Mid-Term Test. Classifying incoming e-mails using the contents, and forwarding them to appropriate departments for further processing. This is referring to §4.6, where AI is used for categorization. RPA are virtual agents that cannot execute physical tasks such as delivery and clearing bins. Sending out individual e-mails with specific information is an application of RPA, but AI is not used. 2. What are the main challenges that driverless vehicles face on a busy urban road? The proximity of neighbouring vehicles. Intermittent mobile Internet connection. Unpredictable human drivers and pedestrians. Finding the way to the destination. Mobile Internet connection should be relatively stable on a busy urban road, and driverless vehicles mostly do not depend on Internet connections. Using Google Map, say, one can easily find the way to the destination. There should be no problem with nearby vehicles, unless they move in unexpected ways, as one can see from the demonstration in §4.3.  3. How are 5G cellular networks useful for robotics? 5G networks support GPS navigation capabilities via Wi-Fi. 5G networks prevent the robot from being taken over by hackers because they allow data to be transmitted only in one direction. 5G networks have low latency, which allows a human to control the robot remotely in almost real time. 5G networks can charge the robot remotely to help it stay powered for longer. This is referring to the part about 5G networks in §4.4. Neither 5G networks nor GPS uses Wi-Fi, and GPS can work independently of 5G networks. 5G networks are not for charging power. Data transmission in 5G networks is not uni-directional. 4. Which of the following is an example of a home appliance that is essentially an AI-powered robot? A human domestic helper that can take care of all the household chores and also the kids’ homework. A cableless machine that can move around and vacuum-clean the floor autonomously and on request. A delivery robot that is able to use lifts, navigate narrow corridors in HDB blocks, and give way to people. A washing machine cum dryer that has a variety of preset programs for the user to choose from. This is referring to Roomba in §4.7. A delivery robot is not a home appliance. A human domestic helper is not a robot. AI is not involved in running the washing machine described.  5. What is the main difference between cobots and usual factory robotics? Cobots generally have a wider range of motion. Cobots are designed to work alongside humans. Cobots can handle more dangerous tasks. Cobots are able to collaborate with one another. This is referring to the part about cobots in §4.5. 6. How could a robot teach itself to walk using Al techniques? It goes through each and every possible way to walk to find the best one. It calculates the best step forward using the physical laws given to it. It learns the most logical way to walk from YouTube videos. It tries and tries, and it is given a reward when it makes forward progress. 7. What can Al do in terms of perception? Listen to and understand the environment; recognize objects/faces. Run simulations to find the best path to travel from A to B. Come up with innovations and invent new software. Make good decisions, e.g., when to trade commodities. 8. Which of the following is the most prominent deficiency of RPA that does not use Al, compared to RPA that uses Al? It is more expensive to develop. It requires data to be in very specific formats. It uses more computing power in general. It needs to learn from a large number of existing examples. S5: Capabilities: Thinking 1. Which is the LEAST feasible way in which one can use AI to help one make better personal purchasing decisions in online marketplaces? Use past purchase records to predict future needs. Analyze customer reviews for a given specific seller. Find the best Internet deal given a specific product. Use past purchase records to recommend products. See §5.8. Purchase records alone may not reflect the needs well, e.g., one may have stocked up when there was promotion, although one may not actually need that much stock. AI-powered RPA from §4.6 can help find the best deal on the Internet, and AI-powered sentiment analysis from §2.2 can help analyze customer reviews. 2. Amongst the following options, which is likely the most pervasive application of AI in education in the future? Open-access e-textbooks. Chatbots as learning assistants. Wikipedia. Generation of images from text prompts. This is referring to the part about the benefits of AI in education in §5.4. AI-powered learning assistants are very useful in personalizing education, providing instant feedback, and making education more widely available. Wikipedia is not (and will probably never be) an application of AI in education. There are limited applications of image generation in education. Open-access e-textbooks are not an application of AI.  3. Which of the following pieces of technology does NOT typically involve AI at present? Machine translation. Open-ended chatbots. Subtitle generators. Scientific calculators. Scientific calculators currently do not use AI. Machine translation is discussed in §2.6. Subtitle generators use speech recognition; cf. §2.7. Open-ended chatbots is discussed in the subsection about chatbots in §2.10. 4. Which of the following is the most direct way in which improvements brought about by AI in weather forecasting may benefit people? They conserve cloud resources. They reverse climate change. They enhance public safety. They help forests grow. Having better weather forecasts help people better prepare for bad weather; see §5.7. Even if there is a way in which better weather forecasts can contribute to the reversing of climate change or the growth of forests, it is almost certain that many other important components need to be involved in the process. It is not clear how better weather forecasts can help conserve cloud resources. 5. Which of the following is NOT an application of AI in banking? Using blockchains to make transactions more secure. Using AI to personalize banking recommendations. Using facial recognition to authenticate transactions. Using chatbots to provide 24/7 customer services. Blockchains are not AI; see §5.10. For chatbots, facial recognition, and personalized recommendations, see §2.10, §3.3, and §5.1 respectively. 6. Which of the following makes best use of digital twins? Finding improvements across operations and testing innovations safely. Detecting fraudulent online transactions that originate from the building. Tracking the movements of people within the building. Helping visitors navigate within the building via AR glasses. 7. How could Al help reduce electricity usage for a data centre? Al can help schedule non-urgent computations at night time when electricity is cheaper. Using temperature sensors at the entrances, Al can detect when cold air leaks out of the centre. By analyzing temperature sensor data, Al can target the cooling at specific computer racks. A network of electric generators can help turn the heat generated by servers into electricity. 8. Which of the following best describes the Fourth Industrial Revolution? Mass digitalization of businesses, including the ways they generate revenue, [...] and store data. Adoption of technology to solve social issues like global warming [...] and socioeconomic disparity. Technological fusion of high-speed digital networks that impacts the physical [...] spheres. Emergence of businesses that offer personalized goods and services tailored to the customer. S6: Tech background 1. Which of the following is an example of unsupervised learning? Learning which e-mails are spam according to what the users report. Learning which news articles fall under the same topic by analyzing the content. Learning how to play football by remembering the behaviours that previously led to goals. Learning how to recognize digits in images using the numerical labels attached to each image. “Learning which news articles fall under the same topic by analyzing the content” involves identifying patterns (i.e., the topics) in unlabelled data (i.e., the news articles); thus it is an instance of unsupervised learning according to the subsection on machine learning in §6.1. User reports and numerical labels make the training data labelled, and so the associated options are not unsupervised. The option about football is guided by rewards (i.e., goals), and so is not unsupervised. (Rather, it falls under reinforcement learning.) 2. What are some advantages of using cloud computing services to support AI in enterprises? Cloud services provide scalable resources that are accessible on demand. Cloud services enable enterprises to provide personalized services to customers. Cloud services are accessible even where there is no Internet connection. Cloud services give enterprises more control over hardware maintenance. This is referring to §6.6. Personalized services can be provided via on-site computing infrastructure as well.  3. Which of the following are the most prominent advantages that the abundance of data gives to Internet giants, such as Meta and Google? The company can take software code straight from this data without having to write their own. The abundance of data allows them to train high-quality AI models and to better understand people. These data are collected into an open archive on the Internet that is freely available to everyone. Once a critical amount of data is reached, AI that surpass human abilities would be born. This is referring to §6.3. As data is of such strategic importance, companies would likely not make them freely available. Although a lot of software code is present in this data, one still has the adapt these to particular applications. The process that goes from large amounts of data to powerful AI is far from automatic: data is not the only bottleneck towards artificial super-intelligence. 4. Which is an advantage of embedding AI into smartphones? This empowers smartphones to take digital photos and videos even when Wi-Fi signals are poor. This allows smartphones to make phone calls and send messages even without network connection. This causes smartphones to consume more power, especially when cloud resources are not available. This enables smartphones to perform voice or image recognition even without network connection. This is referring to §6.7, where we discussed the advantages of edge computing. AI cannot make smartphones make phone calls and send messages when there is no network connection. Taking digital photos and videos do not require AI. Consuming more power is not an advantage.  5. Why can deep neural networks produce answers to questions when these answers are not programmed into them? They are trained using data to solve problems. They use multiple statistical tests to make inferences. They seek optimal solutions using linear programming. They work collaboratively with humans. This is referring to the subsection about machine learning in §6.1. Deep neural networks use neither linear programming nor statistical tests to answer questions. Answers supplied by human collaborators are not actually produced by the deep neural networks. 6. How could one use a GPU to run one's Al models without buying a GPU? Use NVIDIA Tensor Core A100. Use Wireless@SG. Use OpenAl's ChatGPT. Use Google Colab. 7. Why does the development of Al systems often require automatic data extraction from the Internet (i.e., web crawling)? To obtain enough data for training purposes. To monitor which websites are still being maintained. To verify that the approach adopted is not yet obsolete. To enable the Al to query the Internet at runtime.  8. Why is it useful to run Al at the edge? It enables the Al to exploit the latest technological advances. It allows the edge devices to be autonomous. It results in more data being sent to the cloud for processing. It reduces the severity of cybersecurity attacks on the loT.