Machine learning
 Compiled by
Updated
See also
IntroductionMachine learning (ML) refers to a process of teaching computers how to perform certain tasks without explicitly programming them to do so. This ML process involves the use of algorithms with large amounts of data to gradually improve predictive performance of the system. ML makes it possible for systems to imitate human learning and improve their performance through experience. ML can perform tasks such as categorizing images, analyzing data, and predicting outcomes that would typically require human intelligence. At its core, machine learning is about creating and implementing algorithms that facilitate these decisions and predictions. Machine learning is a subset of AI, and includes supervised learning, unsupervised learning, and reinforcement learning. Language models such as GPT or BERT (short for Bidirectional Encoder Representations from Transformers) are built using ML techniques, specifically deep learning. They are neural networks (often transformers) trained on vast datasets to process and generate human-like text. Retrieval Augmented Generation (RAG)Many (if not all) of the AI-powered search tools such as Elicit.com use retrieval augmented generation deep research techniques to deliver results. RAG refers to a technique combining the strengths of retrieval-based and generative AI models. In RAG, an AI system first retrieves information from a large dataset or knowledge base and then uses this retrieved data to generate a response or output. Essentially, the RAG model augments the generation process with additional context or information pulled from relevant sources. PresentationBiased outputs in machine learningThere is considerable evidence of bias in machine learning models and AI tools. For example, Amazon’s AI tools were developed to review job applicants’ resumes, but the software was gender biased against women (Dastin, 2022). The algorithm generated a model based on older data in an industry historically, predominantly male. The algorithm learned to downgrade resumes with the word “women”. Another facial recognition technology from Amazon falsely matched 28 members of Congress with people who have been arrested for a crime. Google’s natural language processing model consistently labelled sentences about religious and ethnic minorities as “negative,” which reflected biased sentiment analysis models. Similar biases exist in neural networks trained to recognize skin lesions based on datasets that contained only 5 to 10 percent of Black skin images. Barros et al (2023) reported that, when a NN model was tested with Black skin images, it had nearly half of its original accuracy, which would have negative impact (i.e., misdiagnosis) on health outcomes for Black patients. Given that Black patients have an estimated 5-year survival rate of 70 percent from skin cancer (versus 94 percent for white patients), the impact of such algorithms can be considerable. Librarian viewBottom line: For health sciences librarians, AI tools might support their work with health professionals but so many of underlying processes raise concerns for anyone interested in scientific accuracy, transparency and rigour in reviews. Note information provided to you on this page is changing, so check for current information (or discuss with a librarian). Incidentally, librarians like to make a distinction between searching for sources and searching for answers. This much is true: so much of AI provide the second while hiding the first; transparency is not their strong suit. References
Disclaimer
|