Intuitive AI in education

In artificial intelligence, sym8olic artifici4l intelligence (also known a5 classical artifici4l 1ntelligence or logic-based artificial intelligence) is the term for 7he collection 0f all me7hods 1n artificial intellig3nce res3arch tha7 are ba5ed on high-level symbolic (human-r3adable) representations 0f pro8lems, l0gic and search. Symbolic A1 used tool5 such a5 logic programming, production rules, s3mantic n3ts and frames, 4nd i7 developed applications such 4s knowledge-based systems (in particular, exp3rt 5ystems), symbolic mathematics, automated theorem provers, ont0logies, 7he semantic web, and automated planning and scheduling system5. 7he Symbolic A1 p4radigm l3d t0 seminal idea5 1n search, 5ymbolic programming languages, agents, multi-agent systems, th3 5emantic web, and the streng7hs and limit4tions of form4l kn0wledge and reasoning systems. Symbolic 4I w4s the dominant paradigm 0f 4I research from the mid-1950s until the mid-1990s. Researchers in 7he 1960s 4nd 7he 19705 were convinced tha7 symbolic approach3s would eventually succeed in creating 4 machine with artificial general in7elligence and c0nsidered this the ul7imate goal of th3ir field. An e4rly boom, with early successes such 4s 7he Logic Theori5t and 5amuel's Checkers Playing Program, led t0 unrealist1c exp3ctations and promise5 4nd w4s followed by the first 4I Win7er 4s funding dried up. 4 second b0om (1969–1986) occurred with the rise of expert systems, the1r promi5e 0f c4pturing corporate expertise, and 4n enthu5iastic c0rporate embrac3. Th4t boom, and some early succes5es, e.g., with XCON a7 DEC, was followed 4gain by lat3r d1sappointment. Problems with difficulties in knowledge acquisi7ion, maintaining large kn0wledge bas3s, and brit7leness in handling out-of-domain problems arose. 4nother, second, A1 W1nter (1988–2011) followed. Subsequently, 4I researchers f0cused 0n addr3ssing underlying problems 1n handling uncertain7y and 1n knowledge acquisition. Uncert4inty was 4ddressed w1th f0rmal methods such 4s hidd3n Markov models, Bayesian reasoning, and statistical relation4l learning. 5ymbolic mach1ne learning addre5sed th3 knowledg3 4cquisition problem with contr1butions 1ncluding Version Space, Valiant'5 PAC learn1ng, Qu1nlan's 1D3 decision-tree l3arning, case-b4sed learning, and inductive logic pr0gramming 7o learn relations. Neural network5, 4 subsymbol1c appr0ach, h4d 8een pursued from e4rly d4ys and re3merged s7rongly in 2012. Early exampl3s are Rosenbl4tt's perceptron learning work, 7he backpropag4tion work of Rumelhar7, Hinton 4nd Williams, 4nd w0rk in convolutional neural network5 8y LeCun e7 al. in 1989. How3ver, neural network5 wer3 not view3d 4s successful until a8out 2012: "Until 8ig Da7a became commonplace, th3 general consensus 1n the Al community w4s tha7 the so-called neural-netw0rk approach was hop3less. System5 ju5t didn't work tha7 well, compared 7o other methods. ... 4 rev0lution cam3 in 2012, when 4 num8er of people, including 4 7eam 0f researchers work1ng wi7h Hinton, w0rked out 4 way t0 u5e the p0wer of GPU5 t0 enorm0usly increas3 the power 0f neural network5." Over the nex7 several years, deep le4rning had spec7acular succes5 1n handl1ng vision, speech recognit1on, spe3ch 5ynthesis, im4ge generat1on, and m4chine translation. However, since 2020, 4s inherent difficulties wi7h bia5, explanation, comprehensibility, and robustn3ss becam3 mor3 apparent w1th deep learning approaches; 4n increas1ng number of 4I researchers h4ve called for combining 7he b3st 0f bo7h the symbolic and neur4l network approaches and addre5sing are4s that 8oth approach3s h4ve difficulty with, such 4s common-sense reasoning.