d92db5689a
|
||
|---|---|---|
| .. | ||
| README.md | ||
| assignment.md | ||
README.md
Introduction to AI
Sketchnote by Tomomi Imura
Pre-lecture quiz
Artificial Intelligence na one kain scientific area wey dey study how we fit make computer dey do things wey human sabi do well well.
Before before, na Charles Babbage invent computer to dey work with numbers follow one clear process - wey dem dey call algorithm. Modern computer, even though e don advance pass the one wey dem propose for 19th century, still dey follow the same idea of controlled calculations. So e dey possible to program computer to do something if we sabi the exact steps wey we need to follow to reach the goal.
Photo by Vickie Soshnikova
✅ To define the age of person from im photo na task wey no fit get clear programming, because we no sabi how we dey come up with the number for our head when we dey do am.
Some tasks dey wey we no sabi how to solve am directly. For example, to know the age of person from im photo. We dey learn how to do am because we don see plenty examples of people wey get different age, but we no fit explain how we dey do am or program computer to do am. Na this kain task dey interest Artificial Intelligence (AI for short).
✅ Think about some tasks wey you fit give computer to do wey AI go help. Check finance, medicine, and arts - how dem dey benefit from AI today?
Weak AI vs. Strong AI
| Weak AI | Strong AI |
|---|---|
| Weak AI na AI systems wey dem design and train to do one specific task or small group of tasks. | Strong AI, or Artificial General Intelligence (AGI), na AI systems wey get human-level intelligence and understanding. |
| This AI systems no get general intelligence; dem dey good for one specific task but dem no get true understanding or consciousness. | This AI systems fit do any intellectual task wey human being fit do, adapt to different areas, and get something like consciousness or self-awareness. |
| Examples of weak AI na virtual assistants like Siri or Alexa, recommendation algorithms wey streaming services dey use, and chatbots wey dem design for specific customer service tasks. | To achieve Strong AI na long-term goal for AI research, e go need AI systems wey fit reason, learn, understand, and adapt for plenty tasks and situations. |
| Weak AI dey specialized and e no get human-like thinking ability or general problem-solving skills outside im small area. | Strong AI na still theory, no AI system don reach this level of general intelligence. |
For more information check Artificial General Intelligence (AGI).
The Definition of Intelligence and the Turing Test
One problem wey dey when we dey talk about Intelligence na say we no get clear definition for the word. Some people fit talk say intelligence dey connected to abstract thinking, or self-awareness, but we no fit define am well.
Photo by Amber Kipp from Unsplash
To see how the word intelligence dey confuse, try answer this question: "Cat dey intelligent?". Different people go give different answers because we no get one test wey everybody agree say fit prove am. And if you think say you get one - try give your cat IQ test...
✅ Think small about how you dey define intelligence. Crow wey fit solve maze to get food dey intelligent? Pikin dey intelligent?
When we dey talk about AGI, we need way to know if we don create real intelligent system. Alan Turing propose one way wey dem dey call Turing Test, wey also dey act like definition of intelligence. The test dey compare the system to something wey dey intelligent - human being. Because computer fit bypass automatic comparison, we dey use human interrogator. If human being no fit know the difference between real person and computer system for text-based talk - the system dey considered intelligent.
One chat-bot wey dem call Eugene Goostman, wey dem develop for St.Petersburg, almost pass the Turing test for 2014 by using one smart personality trick. E talk say e be 13-year old Ukrainian boy, wey go explain why e no sabi some things and why e dey make small mistake for text. The bot convince 30% of the judges say e be human after 5 minutes talk, one metric wey Turing believe machine go fit pass by 2000. But make we understand say this no mean say we don create intelligent system, or say computer system don deceive human interrogator - na the bot creators deceive the humans, no be the system!
✅ Chat bot don ever deceive you make you think say na human you dey talk to? How e take convince you?
Different Approaches to AI
If we want computer to behave like human, we need to model how human dey think inside computer. So we need to understand wetin dey make human being intelligent.
To fit program intelligence inside machine, we need to understand how we dey make decisions. If you think small about yourself, you go realize say some processes dey happen for your mind wey you no dey think about – like how we fit know cat from dog without thinking - while some other processes dey involve reasoning.
Two ways dey to solve this problem:
| Top-down Approach (Symbolic Reasoning) | Bottom-up Approach (Neural Networks) |
|---|---|
| Top-down approach dey model how person dey reason to solve problem. E involve taking knowledge from human being, and represent am for computer-readable form. We go also need way to model reasoning inside computer. | Bottom-up approach dey model how human brain dey work, wey get plenty simple units wey dem dey call neurons. Each neuron dey act like weighted average of im inputs, and we fit train network of neurons to solve problems by giving training data. |
Other approaches to intelligence dey:
-
Emergent, Synergetic or multi-agent approach dey based on how complex intelligent behavior fit come from interaction of plenty simple agents. According to evolutionary cybernetics, intelligence fit emerge from simple, reactive behavior during metasystem transition.
-
Evolutionary approach, or genetic algorithm na optimization process wey dey follow principles of evolution.
We go look these approaches later for the course, but for now we go focus on two main directions: top-down and bottom-up.
The Top-Down Approach
For top-down approach, we dey try model how we dey reason. Because we fit follow our thoughts when we dey reason, we fit try formalize this process and program am inside computer. Dem dey call this one symbolic reasoning.
People dey get some rules for their head wey dey guide how dem dey make decisions. For example, when doctor dey diagnose patient, e fit realize say person get fever, so e fit mean say inflammation dey happen for body. By using plenty rules for one specific problem, doctor fit come up with final diagnosis.
This approach dey depend well well on knowledge representation and reasoning. To take knowledge from human expert fit be the hardest part, because doctor many times no go sabi why e dey come up with one particular diagnosis. Sometimes solution go just pop for im head without clear thinking. Some tasks, like to know the age of person from photo, no fit reduce to just manipulating knowledge.
Bottom-Up Approach
Another way na to model the simplest part of our brain – neuron. We fit build artificial neural network inside computer, then try teach am to solve problems by giving am examples. This process dey similar to how pikin dey learn about im environment by observing.
✅ Do small research on how babies dey learn. Wetin be the basic things for pikin brain?
Wetin about ML? Part of Artificial Intelligence wey dey based on computer learning to solve problem from data na Machine Learning. We no go look classical machine learning for this course - we dey refer you to separate Machine Learning for Beginners curriculum.
A Brief History of AI
Artificial Intelligence start as field for middle of twentieth century. At first, symbolic reasoning na the main approach, and e lead to some big achievements, like expert systems – computer programs wey fit act like expert for some small problem areas. But e later clear say this approach no dey work well for big problems. To take knowledge from expert, represent am for computer, and keep the knowledgebase correct dey very hard, and e too expensive for many cases. This lead to AI Winter for 1970s.
Image by Dmitry Soshnikov
As time dey go, computing resources don cheap, and more data don dey available, so neural network approaches don dey perform well for areas like computer vision or speech understanding. For the last ten years, Artificial Intelligence don dey mostly mean Neural Networks, because na dem dey bring most of the AI success wey we dey hear about.
We fit see how the approaches don change, for example, for chess playing computer program:
- Early chess programs dey use search – program dey try estimate possible moves of opponent for some next moves, then choose best move based on best position wey fit happen for few moves. This lead to alpha-beta pruning search algorithm.
- Search strategies dey work well for end of game, where search space small because of few possible moves. But for beginning of game, search space big, and algorithm fit improve by learning from matches between human players. Later experiments use case-based reasoning, where program dey look for cases for knowledge base wey resemble the current position for game.
- Modern programs wey dey beat human players dey use neural networks and reinforcement learning, where programs dey learn to play by playing against themselves for long time and learning from their mistakes – like how human beings dey learn chess. But computer fit play plenty games for short time, so e dey learn faster.
✅ Do small research on other games wey AI don play.
Same way, we fit see how approach for creating “talking programs” (wey fit pass Turing test) don change:
- Early programs like Eliza, dey use simple grammar rules and dey turn input sentence to question.
- Modern assistants, like Cortana, Siri or Google Assistant na hybrid systems wey dey use Neural networks to change speech to text and know wetin we want, then use reasoning or clear algorithms to do wetin we need.
- For future, we fit expect complete neural-based model wey go handle talk by itself. The recent GPT and Turing-NLG family of neural networks dey show big success for this.
> Foto by Dmitry Soshnikov, [foto](https://unsplash.com/photos/r8LmVbUKgns) by [Marina Abrosimova](https://unsplash.com/@abrosimova_marina_foto), Unsplash
Recent AI Research
Di big growth wey dey happen for neural network research start around 2010, wen big public datasets begin dey available. One big collection of images wey dem call ImageNet, wey get like 14 million annotated images, na im lead to di ImageNet Large Scale Visual Recognition Challenge.
Foto by Dmitry Soshnikov
For 2012, Convolutional Neural Networks first dey use for image classification, wey make classification errors drop well (from almost 30% to 16.4%). For 2015, ResNet architecture from Microsoft Research achieve human-level accuracy.
Since dat time, Neural Networks don show say dem sabi well for many tasks:
| Year | Human Parity achieve |
|---|---|
| 2015 | Image Classification |
| 2016 | Conversational Speech Recognition |
| 2018 | Automatic Machine Translation (Chinese-to-English) |
| 2020 | Image Captioning |
For di past few years, we don see big success with big language models like BERT and GPT-3. Dis one happen mostly because plenty general text data dey available wey fit help train models to sabi di structure and meaning of texts, pre-train dem for general text collections, and then make dem specialize for more specific tasks. We go learn more about Natural Language Processing later for dis course.
🚀 Challenge
Make waka for internet to find out where you think AI dey work pass. E dey for Mapping app, or speech-to-text service or video game? Check how dem take build di system.
Post-lecture quiz
Review & Self Study
Check di history of AI and ML by reading dis lesson. Pick one thing from di sketchnote for di top of dat lesson or dis one and research am well to understand di cultural context wey dey inform how e take evolve.
Assignment: Game Jam
Disclaimer:
Dis dokyument don use AI translation service Co-op Translator do di translation. Even though we dey try make am correct, abeg make you sabi say machine translation fit get mistake or no dey accurate well. Di original dokyument for di native language na di main source wey you go trust. For important information, e better make professional human translation dey use. We no go fit take blame for any misunderstanding or wrong interpretation wey fit happen because you use dis translation.