CogSci_5AI&ML.pdf

Artificial Intelligence & Machine Learning I. Artificial Intelligence formulations II.Machine Learning III.Expert Systems that use Machine Learning A.Diagnosis of medical disorders B. Mortgage wizard IV. ID3 Machine Learning tool to construct decision tree V. Limitations of Machine Learning − CAPTCH...

Artificial Intelligence & Machine Learning I. Artificial Intelligence formulations II.Machine Learning III.Expert Systems that use Machine Learning A.Diagnosis of medical disorders B. Mortgage wizard IV. ID3 Machine Learning tool to construct decision tree V. Limitations of Machine Learning − CAPTCHA Basic Definitions ✧ Artificial Intelligence (AI): tries to design computer models that accomplish the same cognitive tasks that humans do ✧ Machine Learning: a subset of AI that allows computers to “learn” (i.e., progressively improve performance on a specific task) by creating new algorithms to produce a desired output based on structured (or unstructured) data that is provided ✧ Deep Learning: a subset of Machine Learning involving numerous layers of algorithms ➜ Computer does not need to be provided with structured data ✧ Neural Networks: Networks of algorithms that are similar to the neural networks present in the human brain Artificial Intelligence Formulations There are a number of different artificial intelligence formulations Strong AI: General purpose AI − Machines that possess artificial general intelligence (AGI) − Would be just as smart as humans across the board with the ability to understand and learn any task that a human can Applied AI: AGI isn't going to be created any time soon, but machine learning has made it possible for machines to learn how to master complex tasks (expert systems), including − Playing the ancient Chinese board game Go − Identifying human faces − Translating text into practically every language − Spotting tumors − Driving cars − Identifying animals (Friedenberg & Silverman, 2014) Machine Learning Research in AI and Machine Learning boomed in early 21st century due to growth in computing power and availability of large data sets Machine learning is the key to the success of multibillion-dollar corporations such as Google, Facebook, and Amazon Expert Systems Expert systems: Computer programs that are designed to replicate – and improve on – the performance of human experts in specialized domains, such as medical diagnosis and financial services Can start with human experts and write a program that codifies their collective knowledge Machine learning algorithms can also be used to construct a decision tree by analyzing large databases of examples and deriving rules that can then be used to classify new examples − Machine learning is a subfield in expert systems research Expert systems programs are typically applied in narrowly defined domains to solve very determinate problems, such as diagnosing specific medical disorders Ø MYCIN Early expert system developed in 1970s to diagnose infectious diseases Used knowledge base of about 600 heuristic rules about infectious diseases derived from clinical experts and textbooks Accuracy rate of 69%, which turned out to be significantly higher than diagnoses by infectious disease experts Are expert systems now able to accurately diagnose psychological disorders? Deep learning (a subtype of machine learning that we will discuss in greater detail later) has been used to integrate data obtained from multiple imaging modalities, such as fMRI, MRI, and PET, to effectively classify psychological disorders Most studies have focused on diagnosis of dementia and ADHD, probably due to accessibility of large publicly available neuroimaging data sets Accuracy rates above 90% have been achieved in some studies A few of these studies were also able to accurately predict disease trajectory Studies classifying other mental disorders, including schizophrenia, autism, Parkinson’s, depression, substance abuse, and epilepsy, are also accumulating (Durstewitz, Koppe, and Meyer-Lindenberg, 2019) In addition, we’re nearing the point where we may be able to tailor treatment for psychological disorders based on neuroimaging data Ø People whose depression improved most after behavioral activation therapy had greater brain network connectivity between the anterior insular cortex (involved in assigning importance to events) and the middle temporal gyrus (involved in subjective experience of emotion) (Crowther, Smoski, Minkel et al., 2016) Ø Differences in brain structure and neural connectivity among different regions predicted how well CBT reduced symptoms of those with social anxiety disorder ➜ Estimates of treatment outcome were five times more accurate than estimates using a behavioral assessment tool alone (Whitfield-Gabrieli, Ghosh, Nieto-Castanon et al., 2015) Ø Participants with social anxiety disorder were asked to identify letters behind which occasionally lurked pictures of angry faces ➜ Those who struggled most to avoid being distracted by the threatening stimuli—indicated by more activity in dorsal anterior cingulate cortex— showed the most symptom improvement when treated with CBT (Krumpp, Fitzgerald, Piejko et al, 2016) Ø Online mortgage wizards can gauge an applicant’s “mortgage-worthiness” This is represented through decision trees… Decision tree can be translated to explicit rules Ø Ex: IF income between $40K & $75K AND Applicant working for 1-5 years AND Credit not good THEN no loan (Friedenberg & Silverman, 2006) But how would decision tree be designed? 1) Could ask team of mortgage loan officers to work out a decision tree that captures the practices at their bank 2) Alternatively, could write a program that was capable of producing its own decision tree − Would use info from huge database of all loan decisions that the bank has made over a long period of time, together with all the relevant info regarding income, work history, credit rating, etc. − This is classic example of machine learning v Could also use a combination of #1 and #2 ID3: Highly influential machine learning tool that can be used to construct a decision tree Uses a statistical measure of informativeness of each attribute, called information gain − Algorithm basically works out how well each attribute organizes the examples − It then assigns the attribute with the highest information gain to the first node on the tree − Then the process is repeated Ø Ex: We want to decide whether or not the weather is suitable for playing tennis We collect information for two weeks − Each day, log the principal meteorological data − Also note whether or not we decide to play tennis that day Use this information to construct a decision tree that can be used the future Attributes The target attribute: Play Tennis? The other attributes with the values they can take: Outlook? {sunny, overcast, rain} Temperature? {hot, mild, cool} Humidity? {high, low, normal} Wind? {weak, strong} Database Day Outlook? Temperature? Humidity? Wind? Play Tennis? D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 D11 D12 D13 D14 Sunny Sunny Overcast Rain Rain Rain Overcast Sunny Sunny Rain Sunny Overcast Overcast Rain Hot Hot Hot Mild Cool Cool Cool Mild Cool Mild Mild Mild Hot Mild High High High High Normal Normal Normal High Normal Normal Normal High Normal High Weak Strong Weak Weak Weak Strong Strong Weak Weak Weak Strong Strong Weak Strong No No Yes Yes Yes No Yes No Yes Yes Yes Yes Yes No ✧ Calculate correlations between the target attribute and the other attributes ✧ Then find the attribute with the highest information gain Gain (S, Outlook?) Gain (S, Temperature?) Gain (S, Humidity?) Gain (S, Wind?) − The information gain is highest for Outlook? ➜ That is the attribute assigned to the first node in the decision tree = = = = 0.246 0.029 0.151 0.048 Meaning of correlation: Ø Examples – positive or negative correlation? TV watching and obesity ➜ Positive Learning ability and alcohol consumption ➜ Negative Machine learning algorithms such as ID3 are highly dependent upon how their databases are labeled and categorized Ø Ex: There is ID3-based expert system for diagnosing diseases in soybean crops System must be fed detailed information, such as − Attributes of leaves, fruit pods, and seeds − General facts about the weather ☞ That is, system must be fed data that has already been interpreted ➜ Cannot work with raw data such as photographs of diseased soybean plants CAPTCHA ✧ Fact that it is very difficult for a machine to work with raw uninterpreted data, such as an image, forms the basis of Google’s CAPTCHA (Completely Automated Public Turing Test to Keep Computers and Humans Apart) tool for detecting bots on the internet Point was to stop spam on the internet, e.g., preventing scalpers from writing a computer program that buys every ticket in a fraction of a second They work because humans can read the distorted text but computers and bots can’t ✧ However, computer vision technology started getting so good that bots were solving these CAPTCHA and getting through ✧ So engineers tried to make CAPTCHAs more difficult to solve by adding more twists and turns and noise, and throwing in random lines ✧ Unfortunately, Google research found that Humans got these complex CAPTCHAs right only 33% of the time Advanced computer technology at Google was getting them right 99.8% of the time ✧ Enter ReCAPTCHA, which was used in Google’s book digitalization project ReCAPTCHA showed words from books that computers couldn’t understand When enough people on the internet solving these CAPTCHAs wrote the same word for a piece of text shown, that word would be confirmed and uploaded to an ebook database ✧ Latest rendition: No CAPTCHA reCAPTCHA When you click “I’m not a robot”, it sends over an http request to Google with a whole bunch of useful information, like… − Your IP address, your country, a timestamp − Information from your browsers, such as the way you move your cursor just moments before entering the checkbox, how you were scrolling the page before the click, the time interval between different browser events, and other secret variables For a small % of users where risk analysis engine isn’t sure, No CAPTCHA will ask them to complete an additional challenge: an image recognition CAPTCHA, e.g., select all images that contain a part of a car or street signs If you prove you’re a human once this way, Google’s engine will remember, so that next time after clicking the check box, you’ll be able to pass right through Computer-Based Personality Judgments Study found that Machine Learning based judgments of people’s personality using Facebook Likes Were more accurate than those made by the participants’ Facebook friends − Method: Accuracy was assessed by comparison with 100-item personality survey Had higher external validity when predicting life outcomes such as substance use, political attitudes, and physical health than friends had − Method: Participants’ self-rated personality scores, as well as humans’ and computers’ judgments were entered into regression models to predict 13 life outcomes and traits previously shown to be related to personality Even outperformed self-rated personality scores for Facebook activities, substance use, field of study, and network size; and was comparable in predicting political attitudes and social network characteristics (Youyou, Kosinski, & Stillwell, 2015) Companies gather these “big data” for advertisers (who then personalize the ads we see), and for political candidates v Implications for psychological assessment, marketing, and privacy are huge… Video Reference Video excerpted from: I'm not a robot https://www.youtube.com/watch?v=jCr6rNaZ9EU