How our mind learns?

"Five DECADES of research into artificial neural networks have earned Geoffrey Hinton the moniker of the Godfather of artificial intelligence (AI). Work by his group at the University of Toronto laid the foundations for today’s headline-grabbing AI models, including ChatGPT and LaMDA. These can write coherent (if uninspiring) prose, diagnose illnesses from medical scans and navigate self-driving cars. But for Dr Hinton, creating better models was never the end goal. His hope was that by developing artificial neural networks that could learn to solve complex problems, light might be shed on how the brain’s neural networks do the same.

Brains learn by being subtly rewired: some connections between neurons, known as synapses, are strengthened, while others must be weakened.

 But because the brain has billions of neurons, of which millions could be involved in any single task, scientists have puzzled over how it knows which synapses to tweak and by how much. 

Dr Hinton popularised a clever mathematical algorithm known as backpropagation to solve this problem in artificial neural networks. 

But it was long thought to be too unwieldy to have evolved in the human brain. Now, as AI models are beginning to look increasingly human-like in their abilities, scientists are questioning whether the brain might do something similar after all.

Working out how the brain does what it does is no easy feat. Much of what neuroscientists understand about human learning comes from experiments on small slices of brain tissue, or handfuls of neurons in a Petri dish. It’s often not clear whether living, learning brains work by scaled-up versions of these same rules, or if something more sophisticated is taking place. Even with modern experimental techniques, wherein neuroscientists track hundreds of neurons at a time in live animals, it is hard to reverse-engineer what is really going on.

One of the most prominent and longstanding theories of how the brain learns is Hebbian learning. The idea is that neurons which activate at roughly the same time become more strongly connected; often summarised as “cells that fire together wire together”. Hebbian learning can explain how brains learn simple associations—think of Pavlov’s dogs salivating at the sound of a bell. But for more complicated tasks, like learning a language, Hebbian learning seems too inefficient. Even with huge amounts of training, artificial neural networks trained in this way fall well short of human levels of performance.

Today’s top AI models are engineered differently. To understand how they work, imagine an artificial neural network trained to spot birds in images. Such a model would be made up of thousands of synthetic neurons, arranged in layers. Pictures are fed into the first layer of the network, which sends information about the content of each pixel to the next layer through the AI equivalent of synaptic connections. Here, neurons may use this information to pick out lines or edges before sending signals to the next layer, which might pick out eyes or feet. This process continues until the signals reach the final layer responsible for getting the big call right: “bird” or “not bird”.

Integral to this learning process is the so-called backpropagation-of-error algorithm, often known as backprop. If the network is shown an image of a bird but mistakenly concludes that it is not, then—once it realises the gaffe—it generates an error signal. This error signal moves backwards through the network, layer by layer, strengthening or weakening each connection in order to minimise any future errors. If the model is shown a similar image again, the tweaked connections will lead the model to correctly declare: “bird”.

Neuroscientists have always been sceptical that backpropagation could work in the brain. In 1989, shortly after Dr Hinton and his colleagues showed that the algorithm could be used to train layered neural networks, Francis Crick, the Nobel laureate who co-discovered the structure of DNA, published a takedown of the theory in the journal Nature. Neural networks using the backpropagation algorithm were biologically “unrealistic in almost every respect” he said.

For one thing, neurons mostly send information in one direction. For backpropagation to work in the brain, a perfect mirror image of each network of neurons would therefore have to exist in order to send the error signal backwards. In addition, artificial neurons communicate using signals of varying strengths. Biological neurons, for their part, send signals of fixed strengths, which the backprop algorithm is not designed to deal with.

All the same, the success of neural networks has renewed interest in whether some kind of backprop happens in the brain. There have been promising experimental hints it might. A preprint study published in November 2023, for example, found that individual neurons in the brains of mice do seem to be responding to unique error signals, one of the crucial ingredients of backprop-like algorithms long thought lacking in living brains.

Scientists working at the boundary between neuroscience and AI have also shown that small tweaks to backprop can make it more brain-friendly. One influential study showed that the mirror-image network once thought necessary does not have to be an exact replica of the original for learning to take place (albeit more slowly for big networks). This makes it less implausible. Others have found ways of bypassing a mirror network altogether. If artificial neural networks can be given biologically realistic features, such as specialised neurons that can integrate activity and error signals in different parts of the cell, then backprop can occur with a single set of neurons. Some researchers have also made alterations to the backprop algorithm to allow it to process spikes rather than continuous signals.

Other researchers are exploring rather different theories. In a paper published in Nature Neuroscience earlier this year, Yuhang Song and colleagues at Oxford University laid out a method that flips backprop on its head. In conventional backprop, error signals lead to adjustments in the synapses, which in turn cause changes in neuronal activity. The Oxford researchers proposed that the network could change the activity in the neurons first, and only then adjust the synapses to fit. They called this prospective configuration.

When the authors tested out prospective configuration in artificial neural networks they found that they learned in a much more human-like way—more robustly and with less training—than models trained with backprop. They also found that the network offered a much closer match for human behaviour on other very different tasks, such as one that involved learning how to move a joystick in response to different visual cues.

Learning the hard way

For now though, all of these theories are just that. Designing experiments to prove whether backprop, or any other algorithm, is at play in the brain is surprisingly tricky. For Aran Nayebi and colleagues at Stanford University this seemed like a problem AI could solve.

The scientists used one of four different learning algorithms to train over a thousand neural networks to perform a variety of tasks. They then monitored each network during training, recording neuronal activity and the strength of synaptic connections. Dr Nayebi and his colleagues then trained another supervisory meta-model to deduce the learning algorithm from the recordings. They found that the meta-model could tell which of the four algorithms had been used by recording just a couple of hundreds of virtual neurons at various intervals during learning. The researchers hope such a meta-model could do something similar with equivalent recordings of a real brain.

Identifying the algorithm, or algorithms, that the brain uses to learn would be a big step forward for neuroscience. Not only would it shed light on how the body’s most mysterious organ works, it could also help scientists build new AI-powered tools to try to understand specific neural processes. Whether it could lead to better AI algorithms is unclear. For Dr Hinton, at least, backprop is probably superior to whatever happens in the brain." [1]

1.  Great minds. The Economist; London Vol. 452, Iss. 9410,  (Aug 17, 2024): 63, 64.
 

O, ką dirbtinis intelektas gali padaryti

 

 

"Dirbtinį intelektą (AI) galima apibūdinti kaip meną priversti kompiuterius daryti dalykus, kurie žmonėms atrodo protingi. 

 

Šia prasme jis jau yra plačiai paplitęs. Satnav programinė įranga naudoja paieškos algoritmus, kad surastų greičiausią kelią iš jūsų namų į tą naują restoraną; lėktuvai leidžiasi patys, kad atpažintų greitį viršijančio automobilio numerio raides, kai AI veikia nuosekliai ir patikimai, paleidžiant seną pokštą, tai tiesiog vadinama inžinerija (atvirkščiai, yra dar vienas pokštas, AI - tai dar neveikianti medžiaga).

 

 Dirbtinis intelektas, kuris dabar prikausto tiek daug pasaulio dėmesio ir sunaudoja didžiulį kiekį skaičiavimo galios ir elektros energijos, yra pagrįstas metodika, vadinama giliuoju mokymusi. Giluminio mokymosi metu linijinė algebra (konkrečiai, matricos daugyba) ir statistika naudojamos modeliams išgauti ir išmokti iš didelių duomenų rinkinių mokymo proceso metu. Dideli kalbų modeliai (LLM), pvz., Google Gemini arba OpenAI GPT, buvo mokomi naudoti teksto, vaizdų ir vaizdo įrašų gausą ir išugdė daug gebėjimų, įskaitant „atsirandančius“, kuriems jie nebuvo specialiai mokomi (su daug žadančių, bet ir nerimą keliančių pasekmių). Dabar yra daugiau specializuotų, konkrečiai domenui pritaikytų tokių modelių versijų, skirtų vaizdams, muzikai, robotikai, genomikai, medicinai, klimatui, orams, programinės įrangos kodavimui ir kt.

 

 Už žmogaus supratimo ribų

 

 Sparti pažanga šioje srityje paskatino prognozes, kad dirbtinis intelektas „perima vaistų kūrimą“, kad jis „pakeis kiekvieną Holivudo pasakojimo aspektą“ ir kad jis gali „pakeisti patį mokslą“ (visi teiginiai, pateikti praeityje šiame laikraštyje). Teigiama, kad dirbtinis intelektas pagreitins mokslinius atradimus, automatizuos nuobodžius baltųjų apykaklių darbus ir sukurs nuostabių naujovių, kurių dar neįsivaizduojama. Tikimasi, kad dirbtinis intelektas padidins efektyvumą ir paskatins ekonomikos augimą. Tai taip pat gali pakeisti darbo vietas, kelti pavojų privatumui ir saugumui ir sukelti etinių problemų. Tai jau pranoksta žmogaus supratimą apie tai, ką jis daro.

 

 Tyrėjai vis dar aiškinasi, ką dirbtinis intelektas galės ir ko negalės. Iki šiol pasirodė, kad didesni modeliai, paruošti, naudojant daugiau duomenų, yra pajėgesni. Tai paskatino manyti, kad ir toliau pridėjus daugiau, dirbtinis intelektas bus geresnis. Buvo atlikti „mastelio keitimo dėsnių“, rodančių, kaip modelio dydis ir mokymo duomenų apimtis sąveikauja, siekiant pagerinti LLM, tyrimai. Bet kas yra „geresnis“ LLM? Ar jis teisingai atsako į klausimus, ar kyla kūrybinių idėjų?

 

 Taip pat sudėtinga numatyti, kaip gerai esamos sistemos ir procesai galės panaudoti AI. Iki šiol AI galia labiausiai išryškėja atliekant atskiras užduotis. Pateikite riaušių minios atvaizdus, ​​o dirbtinio intelekto modelis, paruoštas šiam konkrečiam tikslui, gali atpažinti minioje esančius veidus valdžiai. Laikykite LLM teisės egzaminą ir jam seksis geriau, nei jūsų vidutinis vidurinės mokyklos mokinys. Tačiau neterminuotų užduočių našumą įvertinti sunkiau.

 

 Šiuolaikiniai dideli dirbtinio intelekto modeliai labai gerai generuoja dalykus, pradedant poezija ir baigiant fotorealistiniais vaizdais, remiantis jų mokymo duomenų šablonais. Tačiau tokie modeliai ne taip gerai sprendžia, kurie iš jų sukurtų dalykų yra prasmingiausi arba tinkamiausi konkrečioje situacijoje. Jiems mažiau sekasi logika ir samprotavimas. Neaišku, ar daugiau duomenų atrakins galimybę nuosekliai samprotauti, ar reikės visiškai skirtingų modelių. Gali būti, kad dirbtinio intelekto ribos ilgą laiką bus tokios, kad jo galiai panaudoti reikės žmonių samprotavimų.

 

 Išsiaiškinti, kokios yra šios ribos, bus svarbu tokiose srityse, kaip sveikatos priežiūra. Tinkamai naudojant AI gali anksčiau aptikti vėžį, išplėsti prieigą prie paslaugų, pagerinti diagnozę ir individualizuoti gydymą. Remiantis metaanalize, paskelbta balandžio mėn. žurnale npj Digital Medicine, dirbtinio intelekto algoritmai gali pranokti žmones gydytojus, atliekant tokias užduotis. Tačiau jų mokymas gali juos suklaidinti tokiais būdais, kurie rodo žmogaus įsikišimo vertę.

 

 Pavyzdžiui, dirbtinio intelekto modeliai yra linkę sustiprinti žmogaus šališkumą dėl „duomenų paskirstymo poslinkių“; Diagnostinis modelis gali klysti, jei jis dažniausiai naudojamas pagal balto žmogaus odos atvaizdus, ​​o tada jam suteikiamas juodaodžio odos vaizdas. AI derinimas su kvalifikuotu žmogumi pasirodė esąs veiksmingiausias. Straipsnyje nustatyta, kad dirbtinį intelektą naudojantys gydytojai sugebėjo padidinti žmonių, kuriems teisingai diagnozuotas vėžys, skaičių nuo 81,1 % iki 86,1 %, o taip pat padidino žmonių, kuriems teisingai pasakė, kad neserga vėžiu, skaičių. Kadangi dirbtinio intelekto modeliai dažniausiai daro kitokias klaidas, nei žmonės, buvo pastebėta, kad dirbtinio intelekto ir žmonių partnerystė pranoksta tiek dirbtinį intelektą, tiek žmones.

 

 Robotinis metodas

 

 Žmonėms gali prireikti mažiau tyrinėti naujas mokslo hipotezes. 2009 m. Rossas Kingas iš Kembridžo universiteto pasakė, kad jo galutinis tikslas buvo sukurti sistemą, kuri veiktų, kaip savarankiška laboratorija arba kaip „mokslininkas robotas“. Dr Kingo AI mokslininkas, vadinamas Adamas, buvo sukurtas taip, kad iškeltų hipotezes, naudotų savo robotinę ranką eksperimentams atlikti, rinkti rezultatus su jutikliais ir juos analizuoti. Skirtingai, nei magistrantūros studentai ir doktorantai, Adomui niekada nereikia daryti pertraukos valgyti ar miegoti. Tačiau tokio tipo AI sistemos (kol kas) apsiriboja gana siauromis sritimis, tokiomis, kaip vaistų atradimas ir medžiagų mokslas. Lieka neaišku, ar jie duos daug daugiau, nei tik padidins žmonių atliekamus tyrimus.

 

 Dirbtinio intelekto metodai moksle naudojami dešimtmečius, siekiant klasifikuoti, atsijoti ir analizuoti duomenis bei daryti prognozes. Pavyzdžiui, projekto CETI mokslininkai surinko didelį banginių vokalizacijos duomenų rinkinį, tada parengė šių duomenų AI modelį, kad išsiaiškintų, kurie garsai gali turėti reikšmę. Arba apsvarstykite AlphaFold, gilųjį neuronų tinklą, kurį sukūrė Google DeepMind. Išmokyta naudoti didžiulę baltymų duomenų bazę, ji gali greitai ir tiksliai numatyti trijų matmenų baltymų formas – užduotį, kuriai kažkada reikėjo kruopštaus žmonių eksperimentavimo ir matavimo dienų. GNoME, kita DeepMind sukurta AI sistema, skirta padėti atrasti naujas medžiagas, turinčias specifinių cheminių savybių.

 

 AI taip pat gali padėti suprasti didelius duomenų srautus, kurie kitu atveju būtų didžiuliai tyrėjams, nesvarbu, ar tai susiję su dalelių greitintuvo rezultatų atsijojimu, siekiant nustatyti naujas subatomines daleles, ar neatsilikti nuo mokslinės literatūros. Jokiam žmogui, kad ir koks išrankus skaitytojas būtų, visiškai neįmanoma suvirškinti kiekvieno mokslinio darbo, kuris gali būti susijęs su jo darbu. Vadinamosios literatūra pagrįstos atradimų sistemos gali analizuoti šiuos teksto kalnus, kad surastų tyrimų spragas, naujais būdais sujungtų senas idėjas ar net pasiūlytų naujas hipotezes. Tačiau sunku nustatyti, ar tokio tipo dirbtinio intelekto darbas bus naudingas. AI gali būti ne ką geresnis už žmones netikėtais dedukciniais šuoliais; Vietoj to jis gali tiesiog pritarti įprastiems, gerai numintiems tyrimų keliams, kurie prie niekur įdomaus veda.

 

 Švietimo srityje nerimaujama, kad dirbtinis intelektas, ypač tokie robotai, kaip „ChatGPT“, iš tikrųjų gali būti kliūtis originaliam mąstymui. Remiantis 2023 m. švietimo bendrovės Chegg atliktu tyrimu, 40 % studentų visame pasaulyje naudojo dirbtinį intelektą mokykliniams darbams, daugiausia rašymui. Tai paskatino kai kuriuos mokytojus, profesorius ir mokyklų rajonus uždrausti AI pokalbių robotus. Daugelis baiminasi, kad jų naudojimas trukdys ugdyti problemų sprendimo ir kritinio mąstymo įgūdžius, nes bus sunku išspręsti problemą ar ginčytis. Kiti mokytojai ėmėsi visiškai kitokios krypties, AI suprato, kaip įrankį, ir įtraukė jį į užduotis. Pavyzdžiui, mokinių gali būti paprašyta naudoti „ChatGPT“, kad parašytų esė tam tikra tema, o tada sukritikuotų, kas joje negerai.

 

 Palaukite, ar pokalbių robotas parašė šią istoriją?

 

 Šiuolaikinis generatyvus AI ne tik sukuria tekstą vienu mygtuko paspaudimu, bet ir per kelias sekundes gali sukurti vaizdus, ​​​​garsą ir video. Tai gali supurtyti žiniasklaidos verslo reikalus nuo podcast'ų iki vaizdo žaidimų iki reklamos. Dirbtinio intelekto įrankiai gali supaprastinti redagavimą, sutaupyti laiko ir sumažinti kliūtis patekti į rinką. Tačiau dirbtinio intelekto sukurtas turinys gali kelti pavojų kai kuriems menininkams, pvz., iliustratoriams ar balso aktoriams. Laikui bėgant, gali būti įmanoma sukurti ištisus filmus, naudojant AI varomus žmonių aktorių simuliakrus arba visiškai dirbtinius.

 

 Vis dėlto dirbtinio intelekto modeliai patys negali nei sukurti, nei išspręsti problemų (arba vis dar ne). Jie yra tik sudėtingos programinės įrangos dalys, o ne jautrios ar savarankiškos. Jie pasikliauja žmonių naudotojais, kurie juos iškvies ir paragins, o tada pritaikys arba atmes rezultatus. Dirbtinio intelekto revoliucinis pajėgumas, tiek geresnis, tiek blogesnis, vis dar priklauso nuo žmonių ir žmogaus sprendimo." [1]

 

1. Oh, the things AI can do. The Economist; London Vol. 452, Iss. 9410,  (Aug 17, 2024): 55, 56.

Oh, the things AI can do

 

"Artificial Intelligence (AI) can be described as the art of getting computers to do things that seem smart to humans. In this sense it is already pervasive. Satnav software uses search algorithms to find the quickest route from your house to that new restaurant; airplanes land themselves; traffic cameras use optical character recognition to identify the letters on the number plate of a speeding car; thermostats adjust their temperature settings based on who is at home. This is all AI, even if it is not marketed as such. When AI works consistently and reliably, runs an old joke, it is just called engineering. (Conversely AI, goes another joke, is the stuff that does not quite work yet.)

The AI that is hogging so much of the world’s attention now—and sucking up huge amounts of computing power and electricity—is based on a technique called deep learning. In deep learning linear algebra (specifically, matrix multiplications) and statistics are used to extract, and thus learn, patterns from large datasets during the training process. Large language models (LLMs) like Google’s Gemini or OpenAI’s GPT have been trained on troves of text, images and video and have developed many abilities, including “emergent” ones they were not explicitly trained for (with promising implications, but also worrying ones). More specialised, domain-specific versions of such models now exist for images, music, robotics, genomics, medicine, climate, weather, software-coding and more.

Beyond human comprehension

Rapid progress in the field has led to predictions that AI is “taking over drug development”, that it will “transform every aspect of Hollywood storytelling”, and that it might “transform science itself” (all claims made in this newspaper within the past year). It is said that AI will speed up scientific discovery, automate away the tedium of white-collar jobs and lead to wondrous innovations not yet imaginable. AI is expected to improve efficiency and drive economic growth. It might also displace jobs, endanger privacy and security, and lead to ethical conundrums. It has already outrun human understanding of what it is doing.

Researchers are still getting a handle on what AI will and will not be able to do. So far, bigger models, trained on more data, have proved more capable. This has encouraged a belief that continuing to add more will make for better AI. Research has been done on “scaling laws” that show how model size and the volume of training data interact to improve LLMs. But what is a “better” LLM? Is it one that correctly answers questions, or that comes up with creative ideas?

It is also tricky to predict how well existing systems and processes will be able to make use of AI. So far, the power of AI is most apparent in discrete tasks. Give it images of a rioting mob, and an AI model, trained for this specific purpose, can identify faces in the crowd for the authorities. Give an LLM a law exam, and it will do better than your average high-schooler. But performance on open-ended tasks is harder to evaluate.

The big AI models of the moment are very good at generating things, from poetry to photorealistic images, based on patterns represented in their training data. But such models are less good at deciding which of the things they have generated make the most sense or are the most appropriate in a given situation. They are less good at logic and reasoning. It is unclear whether more data will unlock the capability to reason consistently, or whether entirely different sorts of models will be needed. It is possible that for a long time the limits of AI will be such that the reasoning of humans will be required to harness its power.

Working out what these limits are will matter in areas like health care. Used properly, AI can catch cancer earlier, expand access to services, improve diagnosis and personalise treatment. AI algorithms can outperform human clinicians at such tasks, according to a meta-analysis published in April in npj Digital Medicine. But their training can lead them astray in ways that suggest the value of human intervention.

For example, AI models are prone to exacerbating human bias due to “data distribution shifts”; a diagnostic model may make mistakes if it is trained mostly on images of white people’s skin, and then given an image of a black person’s skin. Combining AI with a qualified human proved the most effective. The paper showed that clinicians using AI were able to increase the share of people they correctly diagnosed with cancer from 81.1% to 86.1%, while also increasing the share of people told correctly they were cancer-free. Because AI models tend to make different mistakes from humans, AI-human partnerships have been seen to outperform both AI and humans alone.

The robotic method

Humans might be less necessary to explore new hypotheses in science. In 2009 Ross King at the University of Cambridge said that his ultimate goal was to design a system that will function as an autonomous lab, or as a “robot scientist”. Dr King’s AI scientist, called Adam, was engineered to come up with hypotheses, use its robotic arm to perform experiments, collect results with its sensors and analyse them. Unlike graduate students and postdocs, Adam never needs to take a break to eat or sleep. But AI systems of this type are (for now) restricted to relatively narrow domains such as drug discovery and materials science. It remains unclear whether they will deliver much more than incremental gains over human-led research.

AI techniques have been used in science for decades, to classify, sift and analyse data, and to make predictions. For example, researchers at Project CETI collected a large dataset of whale vocalisations, then trained an AI model on this data to work out which sounds might have meaning. Or consider AlphaFold, a deep neural network developed by Google DeepMind. Trained on a massive protein database, it can quickly and accurately predict the three-dimensional shapes of proteins, a task that once required days of careful experimentation and measurement by humans. GNoME, another AI system developed by DeepMind, is intended to assist in the discovery of new materials with specific chemical properties (see diagram).

AI can also help make sense of large flows of data that would otherwise be overwhelming for researchers, whether that involves sifting through results from a particle collider to identify new subatomic particles, or keeping up with scientific literature. It is quite impossible for any human, no matter how fastidious a reader, to digest every scientific paper that might be relevant to their work. So-called literature-based discovery systems can analyse these mountains of text to find gaps in research, to combine old ideas in novel ways or even to suggest new hypotheses. It is difficult to determine, though, whether this type of AI work will prove beneficial. AI may not be any better than humans at making unexpected deductive leaps; it may instead simply favour conventional, well-trodden paths of research that lead nowhere exciting.

In education there are concerns that AI—and in particular bots like ChatGPT—might actually be an impediment to original thinking. According to a study done in 2023 by Chegg, an education company, 40% of students around the world used AI to do their school work, mostly for writing. This has led some teachers, professors and school districts to ban AI chatbots. Many fear that their use will interfere with the development of problem-solving and critical-thinking skills through struggling to solve a problem or make an argument. Other teachers have taken an altogether different tack, embracing AI as a tool and incorporating it into assignments. For example, students might be asked to use ChatGPT to write an essay on a topic and then critique it on what it gets wrong.

Wait, did a chatbot write this story?

As well as producing text at the click of a button, today’s generative AI can produce images, audio and videos in a matter of seconds. This has the potential to shake things up in the media business, in fields from podcasting to video games to advertising. AI-powered tools can simplify editing, save time and lower barriers to entry. But AI-generated content may put some artists, such as illustrators or voice actors, at risk. In time, it may be possible to make entire films using AI-driven simulacra of human actors—or entirely artificial ones.

Still, AI models can neither create nor solve problems on their own (or not yet anyway). They are merely elaborate pieces of software, not sentient or autonomous. They rely on human users to invoke them and prompt them, and then to apply or discard the results. AI’s revolutionary capacity, for better or worse, still depends on humans and human judgment." [1]

1. Oh, the things AI can do. The Economist; London Vol. 452, Iss. 9410,  (Aug 17, 2024): 55, 56.