Deep Learning A to Z · 3-1

How AI Works — One Stimulus, Four Paths

A dark shadow approaches. The human brain, machine learning, deep learning, and the fruit fly brain each process this scene along a different path. We run the four paths side by side, then pit machine learning, where people choose the features, against deep learning, which learns even the features by itself, on the same scenes. Finally, we separate the cases where computations resembling neural networks have been confirmed in real brains from the parts that remain estimates.

Written by Sangsun Ahn (CEO, M-Robo Inc.) · About the author

About 86 billionNeurons in the human brain [R1]
139,255Neurons in the complete wiring diagram of the fruit fly brain [R2]
1959Neurons responding to orientation reported in the cat visual cortex [R3]
4 pathsAlong which the same stimulus flows at once
5 stepsHow humans and AI learn
2Code examples you can run right away in Colab
1Four paths

Four Paths — The Same Stimulus Flows at Once

A single stimulus enters four systems at the same time. They differ in the number of stages, what each stage does, and who decided what it does.

Stimulus
Path to inspect

Explained for beginners

All four paths share the same main flow: take in → organize → judge → act. The difference is "who decided what to use as clues (features)." In machine learning, people decide (yellow border); in deep learning, they are learned from data (green border). The fruit fly's looming-detection circuit is not learned but formed during development (pink border). You could say evolution "decided" it over a long time.

The human brain row shows only the main flow found in textbooks. The real brain has many feedback connections running backward, as well as fast shortcuts (such as the path from the eye to the superior colliculus in the midbrain). For the final "judge" stage, all that is known is that several regions work together, so it is drawn with dashed lines. The speed of the animation has nothing to do with actual processing time.

2Human brain

The Human Brain — Separating What Is Known from What Is Estimated

Each sense has its own receiving organ and its own route. Pick a sense, follow the path its signal takes, and check how firmly each stage is established.

Sense
What is known
  • Each sense has a different receiving organ and first pathway. Vision and hearing pass through the thalamus to the cerebral cortex. Smell skips the thalamus and goes from the olfactory bulb straight to the cortex [R10].
  • Neurons in the primary visual cortex respond strongly to bars or edges tilted at a particular orientation. This was first measured in cats [R3].
  • The monkey inferior temporal cortex has groups of neurons that respond selectively to faces [R4].
  • Learning changes the strength of synaptic connections.
Still estimated or under research
  • How the brain binds color, shape, motion, and sound, processed separately in different regions, into "a single object."
  • Exactly where, and by what computation, the decision "let's avoid it" is made. All that is known is that several regions are involved together.
  • Whether the brain learns by sending errors backward layer by layer, like backpropagation in deep learning. Hypotheses that the brain could do something similar by other means are being studied [R13].
  • The complete wiring diagram of the human brain. The fruit fly's whole-brain wiring diagram has been completed [R2], but the human brain has about 620,000 times as many neurons alone.

Why so many "estimates" for the human brain?

To measure the signals of individual neurons, electrodes must be placed in the brain. In humans this is possible only in very limited cases, such as patients who need surgery anyway. So much of our cell-level knowledge comes from animal experiments on cats, monkeys, and the like, while in humans we rely mainly on methods that measure activity over wide areas, such as brain imaging. In fruit flies, genetic tools can switch only the desired neurons on and off, so "this neuron is necessary for this behavior" can be tested directly. That is why this series uses the fruit fly as its guide.

3Machine learning

Machine Learning — People Design the Features

People: choose features → Machine: learn weightsSamuel, 1959

What people do

Looking at two 12×12 black-and-white scenes (before and after), the system must decide "is an object approaching?" First, people choose the clues for the decision, that is, the features. For example: "how much did the dark area grow?" and "how far did the shadow's center move?" The 288 pixels are reduced to 2 numbers.

What the machine does

It learns a weight for each of the 2 numbers plus a bias, 3 values in all, from 240 training examples (logistic regression). The line on the plane on the right is the result. After training, its skill is measured on 160 test scenes it has never seen.

Approaching (must avoid)Passing byRecedingLights dimmingStillYellow border = wrong prediction
Example scene
Human-chosen features

Explained for beginners

Samuel's 1959 checkers program, which popularized the term "machine learning," worked this way too. People supplied a list of judgment criteria that "seemed relevant to the game," and the program, game after game, learned only how much weight to give each criterion [R11]. People decide what to look at; the machine decides how much to trust it.

So performance depends heavily on choosing good features. With feature A, a scene where "the whole room gets darker" is mistaken for an approaching object. This is not fixed even if such scenes are in the training data, because the information needed to tell them apart from approaching objects is already lost the moment everything is reduced to 2 numbers. It is solved only when a person imagines that situation and designs a new clue, as with feature B. This work of devising features is called feature engineering.

4Deep learning

Deep Learning — Learning Even the Features

288 pixels → 6 hidden neurons → probability of approach1,741 weights

What people do

People do not choose the features. Instead, they decide the shape of the neural network (how many hidden neurons, how many layers) and choose which scenes to show it, that is, the training data.

What the machine does

It takes the 288 pixels as they are. Each of the 6 hidden neurons learns, through its weights, "which cells changing in which way should turn it on." The small grid images in the middle figure are the result: features the neural network made by itself.

Positive weight (on when bright)Negative weight (on when dark)Training loss
Training data
Example scene

Explained for beginners

Deep learning finds features in the data instead of having people choose them. So it can pick up clues people never thought of, and it is especially strong on data such as photos, sounds, and text, where features are hard to put into words. The neural network on this page is a very small model with just one hidden layer, but the principle is the same as in models with dozens of layers. As layers get deeper, the early layers learn simple features such as edges, and the later layers learn complex features that combine them [R12].

It is not free. Machine learning learns 3 weights; deep learning learns 1,741. The more there is to learn, the more examples are needed. When we actually ran Code 2, with only 8 training scenes the feature B approach reached 100% test accuracy while the pixel approach reached 93.9%; only at 240 scenes did the two become equal. And as "Without lighting changes" shows, deep learning also gets things wrong in situations that were not in the data. The human role of choosing features simply shifts to the role of choosing data.

5Confirmed cases

Confirmed Cases — Where Computations Resembling Neural Networks Were Measured in the Brain

We look at four places where experiments have tested how far the claim "neural networks resemble the brain" holds. The figures are illustrative reproductions that convey only the direction of the papers' results; they are not actual measurements.

CaseConfirmed by experimentSimilar thing in neural networksAnalogy strength
Fruit fly LPLC2Responds strongly only to motion spreading outward in all directions, and not to receding or to dimming without motion. Provides input to the escape pathway [R6]A feature detector that turns on only for a specific pattern. The circuit does what feature B in Section 3 was deliberately designed to doDirect Even the computation (outer excitation + inner inhibition) has been worked out
Fruit fly mushroom bodyKenyon cells receive inputs from several glomeruli in nearly random combinations [R7]. Only a few turn on for any one odor [R9], and learning happens at the output-side connections [R8]A wide hidden layer fixed at random + only the output layer is trained. The "fruit fly method" in 3-2, Section 6Direct The structure corresponds. However, in deep learning all hidden layers are usually trained
Cat primary visual cortexNeurons that respond strongly to bars tilted at a specific orientation [R3]Orientation-specific edge features learned by the first layer of image neural networks [R12]Partial The results look alike, but the way they come about differs
Monkey inferior temporal cortexGroups of neurons that respond selectively to faces [R4]. The upper layers of neural networks trained to be good at object recognition predict this region's responses well [R5]Increasingly complex features layer by layer (3-2, Section 4: representation learning)Partial Good prediction, but not evidence of the same computation
6Steps of learning

Five Steps of Learning — Humans and AI in Two Rows

Data → processing → learning → prediction → repetition. We place side by side how a child comes to tell dogs from cats and how an AI learns the same thing from photos.

① Data→ ② Processing→ ③ Learning→ ④ Prediction→ ⑤ Repetition
StepHuman (child)AI (photo classification model)Key difference

Colab Code — Run This Page Yourself

Both code examples use only libraries preinstalled in Colab. After "Download notebook (.ipynb)", open it in Colab with File → Upload notebook, or copy the code and paste it into a new notebook cell.

Code 1 · Human-made features vs self-learned featuresLinked to Sections 3 & 4 · numpy, matplotlib · about 5 s on CPU

This reproduces the Section 3 & 4 experiments on this page. On the same scenes it compares logistic regression with features A and B against a neural network that takes raw pixels (trained on data with / without lighting changes), and it plots the features the neural network made by itself. To use your own data, see the parts marked [To use your own data] in the code.

Results (run on 2026-10-03, numpy 2.4 · matplotlib 3.10 · CPU). The random seed is fixed, so you get the same results. The conclusion matches the browser (feature A and the network "trained without lighting changes" get every lighting scene wrong, and everything else gets them all right). However, because the random number generators differ, the scenes differ slightly, so feature A's accuracy is a little different: 86.9% in the browser (also missing 1 approaching scene) and 87.5% in Colab.
Code 2 · "How many examples are needed?" with scikit-learnLinked to Sections 4 & 6 · numpy, scikit-learn, matplotlib · about 20 s on CPU

This runs the same comparison with libraries used in practice. Increasing the number of training examples from 8 to 240, it plots the test accuracy of human-made features (machine learning) and raw pixels (deep learning). For each size, five samples were drawn and averaged.

Results (run on 2026-10-03, numpy 2.4 · scikit-learn 1.8 · matplotlib 3.10 · CPU). If people already know good features, fewer examples are enough; learning features from pixels needs more examples. Numbers may differ slightly with other versions.

AI Prompts — Using This Code as a Starting Point

These are prompts to paste into AI assistants such as ChatGPT, Qwen, or Claude. Pressing "Copy" attaches the chosen code after the prompt. Replace the parts in 〔 〕 with your own situation.

Code to attach

An AI's answer is only a starting point. Always check any papers, figures, or library functions it suggests against the original sources and official documentation. If you used AI help for a competition entry or paper, disclose the fact and extent of its use according to the rules of the organizers and the journal.

Glossary

This is the same dictionary as the yellow dotted underlines in the text. The same terms are used throughout the series. Click a section name to jump to where the term first appears.

Analogies — and Where They Break Down

Analogies are a first step to understanding, but they can also seed misunderstanding. Fruit fly and brain analogies are marked by strength. Direct evidence that they actually do the same thing · Partial only some properties are similar · None no known counterpart

ConceptEveryday analogyFruit fly / brain analogyWhere the analogy breaks down
FeatureThe clues you check when picking fruit (color, weight, smell)Direct LPLC2 detects just one feature: "motion spreading outward in all directions"Most features a neural network learns are not clues people can put a name to
Machine learningA new hire whose checklist items are set by a senior colleague, and who learns only the score for each item through experiencePartial An insect brain in which only some connections change through learning, on top of detection circuits fixed in advance by evolutionEvolution does not "design" features with a purpose. It is simply the result of selection over long ages
Deep learningAn apprentice who sets their own standards by seeing many cases without a checklistPartial Responses resemble the layered structure of the monkey visual pathwaySimilar responses do not mean it learns the same way. Whether the brain does backpropagation has not been confirmed
Stacked layersA factory line that assembles parts step by stepPartial Primary visual cortex → V2 → V4 → inferior temporal cortexThe brain has many backward feedback connections, so signals do not flow in only one direction
Random expansion + output learningAssigning lockers at random, then only putting new name tags on themDirect Mushroom body: nearly random inputs, only a few cells on, learning on the output sideThe mushroom body's thousands of cells and the details of its learning signal (dopamine) are far more complex than a deep learning model
PreprocessingPrepping ingredients before cookingPartial The retina and lamina tidy up brightness changes firstBiological preprocessing is not a rule set by people but the structure of the sense organs itself
LearningPracticing until it becomes second naturePartial Changes in synaptic connection strengthThe rules differ: by what criterion, which synapses, and by how much they change
Situations not in the dataA test question of a type not in the textbookNoneA person facing a new situation can stop and say "I don't know," but a neural network may give a confident wrong answer (Section 4)
Probability outputThe chance of rain in a weather forecastNoneA high probability from the model does not mean it is right. Outside the training range, the probability cannot be trusted either

About the Author and References

E-1 About the author · E-2 Academic references · E-3 Figure sources

E-1 About the author

Sangsun Ahn (CEO, M-Robo Inc.)

After working at the Korea Electrical Safety Corporation and the Maeil Business Newspaper, he founded M-Robo Inc., a spin-off venture of Maeil Business, where he currently serves as CEO. M-Robo Inc. began as an in-house venture of a media company; it provided a robot journalist (StockBot) service to Kukmin Ilbo and has carried out many projects for the Korea Internet & Security Agency, the Korea Press Foundation, the Korea Communications Agency, and others. It now runs its own AI news service, Money-Robo. As CEO he won the Asia Open Data Hackathon and twice received the Minister of the Interior and Safety Award for contributions to data entrepreneurship, and he is the inventor and registrant of 5 patents in artificial intelligence and blockchain. He is the author of “AI Crime and Self-Defense”, and has developed e-learning content on fintech, statistics, and artificial intelligence for the Online Lifelong Education Institute of Korea University of Technology and Education. He holds a PhD in economics (Dongguk University, micro-mathematical economics). He teaches data and AI courses at several universities, including Kookmin University and Hoseo University. Full profile

Courses

  • Ahn, Sangsun (2025). “Data Science for Business” lecture notes. Department of AI Convergence, Open Cyber University of Korea.
  • Ahn, Sangsun (2025). [K-Digital] AI Natural Language Processing. STEP (Smart Training Education Platform), Online Lifelong Education Institute, Korea University of Technology and Education.
  • Ahn, Sangsun (2024). [K-Digital] Fundamentals of Fintech Data Analysis. STEP (Smart Training Education Platform), Online Lifelong Education Institute, Korea University of Technology and Education.

STEP course links: link.step.or.kr/1qyVoJ · link.step.or.kr/M9na9N · link.step.or.kr/brqEyc
If the links do not work due to site reorganization, please visit www.step.or.kr.

Papers and presentations

  • Ahn, Sangsun (2025). A study on household expenditure prediction models using the Korean Retirement and Income Study panel data: An AI model-based approach. “Future Society”, 16(2). Full text
  • Ahn, Sangsun (2025). A study on household expenditure prediction models using the Korean Retirement and Income Study panel data: An AI approach. Presentation at the KReIS Panel Conference (2025.05.13), National Pension Research Institute, National Pension Service. Slides
  • Ahn, Sangsun (2025). A study on an algorithm for detecting suspected drug transactions from SNS text: Focusing on pattern and context analysis. “Future Society”, 16(1). Full text
  • Ahn, Sangsun (2024). Statistics-based countermeasures against adversarial data for securing anomaly detection models in finance: Empirical analysis and simulation results. “Future Society”, 15(3). Full text
  • Ahn, Sangsun (2024). An AI-based solution for detecting SNS posts based on virtual character traits. Poster presentation, 2024 Summer Newmedia Party (2024.06.19), Seoul Media Institute of Technology.
  • Ahn, Sangsun (2024). Machine learning models using industrial accident panel data. Poster presentation, 10th Panel Study of Workers' Compensation Insurance Conference (2024.05.17), Labor Welfare Research Institute, Korea Workers' Compensation & Welfare Service.
  • Ahn, Sangsun (2024). Analysis and prediction of virtual asset time series data: Comparing time series models and AI models, and suggesting improvements. “Future Society”, 15(1). Full text

E-2 Academic references

  1. [R1] Azevedo, F. A. C., Carvalho, L. R. B., Grinberg, L. T., Farfel, J. M., Ferretti, R. E. L., Leite, R. E. P., Jacob Filho, W., Lent, R., & Herculano-Houzel, S. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. Journal of Comparative Neurology, 513(5), 532–541. doi:10.1002/cne.21974 — Hero, Section 2 (average of about 86.1 billion in adult male brains)
  2. [R2] Dorkenwald, S., Matsliah, A., Sterling, A. R., … Seung, H. S., Murthy, M., & the FlyWire Consortium (2024). Neuronal wiring diagram of an adult brain. Nature, 634(8032), 124–138. doi:10.1038/s41586-024-07558-y — Hero, Section 2 (139,255 neurons, about 50 million chemical synapses)
  3. [R3] Hubel, D. H., & Wiesel, T. N. (1959). Receptive fields of single neurones in the cat's striate cortex. The Journal of Physiology, 148(3), 574–591. doi:10.1113/jphysiol.1959.sp006308 — Hero, Sections 2 & 5
  4. [R4] Desimone, R., Albright, T. D., Gross, C. G., & Bruce, C. (1984). Stimulus-selective properties of inferior temporal neurons in the macaque. Journal of Neuroscience, 4(8), 2051–2062. doi:10.1523/JNEUROSCI.04-08-02051.1984 — Sections 2 & 5
  5. [R5] Yamins, D. L. K., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D., & DiCarlo, J. J. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. PNAS, 111(23), 8619–8624. doi:10.1073/pnas.1403112111 — Section 5
  6. [R6] Klapoetke, N. C., Nern, A., Peek, M. Y., Rogers, E. M., Breads, P., Rubin, G. M., Reiser, M. B., & Card, G. M. (2017). Ultra-selective looming detection from radial motion opponency. Nature, 551(7679), 237–241. doi:10.1038/nature24626 — Sections 1, 3 & 5
  7. [R7] Caron, S. J. C., Ruta, V., Abbott, L. F., & Axel, R. (2013). Random convergence of olfactory inputs in the Drosophila mushroom body. Nature, 497(7447), 113–117. doi:10.1038/nature12063 — Sections 1 & 5
  8. [R8] Aso, Y., Hattori, D., Yu, Y., Johnston, R. M., Iyer, N. A., Ngo, T.-T. B., Dionne, H., Abbott, L. F., Axel, R., Tanimoto, H., & Rubin, G. M. (2014). The neuronal architecture of the mushroom body provides a logic for associative learning. eLife, 3, e04577. doi:10.7554/eLife.04577 — Sections 1 & 5
  9. [R9] Dasgupta, S., Stevens, C. F., & Navlakha, S. (2017). A neural algorithm for a fundamental computing problem. Science, 358(6364), 793–796. bioRxiv preprint — Section 5 (interprets as a computing algorithm the structure in which about 5% of Kenyon cells turn on for a single odor)
  10. [R10] Sela, L., Sacher, Y., Serfaty, C., Yeshurun, Y., Soroker, N., & Sobel, N. (2009). Spared and impaired olfactory abilities after thalamic lesions. Journal of Neuroscience, 29(39), 12059–12069. doi:10.1523/JNEUROSCI.2114-09.2009 — Sections 1 & 2 (smell reaches the cortex without a thalamic relay)
  11. [R11] Samuel, A. L. (1959). Some studies in machine learning using the game of checkers. IBM Journal of Research and Development, 3(3), 210–229. doi:10.1147/rd.33.0210 — Section 3. The definition often quoted as coming from this paper, "the ability to learn without being explicitly programmed," could not be found in the original text, so it is not used here
  12. [R12] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. doi:10.1038/nature14539 — Sections 4 & 5
  13. [R13] Lillicrap, T. P., Santoro, A., Marris, L., Akerman, C. J., & Hinton, G. (2020). Backpropagation and the brain. Nature Reviews Neuroscience, 21(6), 335–346. doi:10.1038/s41583-020-0277-3 — Sections 2 & 6

E-3 Figure sources

  • Brain pathway diagrams in Sections 1 & 2 — schematics conveying only spatial relationships, based on the E-2 sources and general neuroanatomy. Not actual anatomical data
  • Response figures in Section 5 — illustrative reproductions conveying only the direction of the results in [R3], [R4], [R6], and [R7]. Not actual measurements

Tips for Educators

A guide for instructors, on-the-job (OJT) trainers, and tutors. Times are estimates; adjust them to your actual class.

AudienceUniversity students, working adults, and high school students new to AI. As the first part of the series, it can be read without equations. With basic Python, learners can also do the Colab exercises
PrerequisitesNone. It helps to know something like "neurons send and receive electrical signals"
Learning objectives① Compare and explain the paths by which the human brain, machine learning, deep learning, and the fruit fly brain process a stimulus ② Explain the difference between machine learning and deep learning in terms of "who decides the features," and give an example of each one's limits ③ Distinguish what brain research has established from what is estimated
Estimated timeReading 30–40 min · animation demo 20 min · Colab exercises 25 min (Code 1: 10 min, Code 2: 15 min)

Three lesson plans

Lecture, 50 min

Concept-focused

Intro 5 min (Section 1, approaching shadow) → Section 2, four senses 10 min → Sections 3 & 4, lighting-trap showdown 20 min → Section 5, pick one case 10 min → check questions 5 min.

Workplace OJT, 90 min

Focused on applying to work

Sections 3 & 4, 15 min → discussion "What features could people define in our work data?" 20 min → Code 2 exercise 25 min → use the Section 6 table to summarize "five steps of AI adoption for our team" 20 min → prompt cards 10 min.

1:1 tutoring

Pace by level

Beginners: Sections 1, 3, 4, and 6 only. Intermediate learners: have them explain with the plane figure why feature A in Section 3 is not fixed by adding more data, then have them design one more feature themselves in Code 1.

Teaching points by section

SectionQuestions to pause and askHands-on tips
1 Four paths"What do all four paths have in common? Who decided how it is done?"Click each option under "Path to inspect" and compare the border colors (yellow, green, pink)
2 Human brain"What is different about smell alone?"Look at vision, then switch to smell and point out that it skips the thalamus
3 Machine learning"Why is the whole room getting darker mistaken for something approaching?"Select "Lights dimming" with feature A, then switch to B and watch the yellow dots move on the plane
4 Deep learning"If people don't choose the features, what do they have to decide?"Train with "Without lighting changes" to show the failure first, then retrain "With" lighting changes
5 Confirmed cases"How is 'the brain and neural networks are similar' different from 'they are the same'?"Read the analogy strength (Direct / Partial) in the table together with the evidence
6 Steps of learning"Does AI keep learning after it is deployed?"Click "Next step" through to ⑤ Repetition and read the table's "Key difference" aloud

Top 5 common misconceptions

MisconceptionCorrect explanationScene to show
Deep learning is a direct copy of the human brainIt borrowed only the skeleton: layers and connection strengths. Its way of learning (backpropagation) has not been confirmed in the brainSection 2 "estimated" list, Section 5 table
Deep learning is always better than machine learningWhen good features are known and data is scarce, machine learning can be betterCode 2 results in the Section 4 explanation box
Deep learning needs no human input at allInstead of features, choosing the data becomes the human's job. It fails in situations not in the dataSection 4 "Without lighting changes"
Machine learning and deep learning are different fieldsDeep learning is a branch of machine learning: machine learning that learns even the features from dataThe two-row comparison in Section 1
How the brain works has been fully figured outThe main routes of the sensory pathways are known, but decision-making and learning rules are still under researchThe dashed stages in Section 2

Check questions (click to expand the explanation)

Q1. Using the word "feature," distinguish machine learning and deep learning in one sentence each.

In (traditional) machine learning, people choose the features and the machine learns a weight for each feature. In deep learning, the machine learns even the features from the raw data. Section 3 learned 3 weights; Section 4 learned 1,741.

Q2. Why does feature A in Section 3 get "Lights dimming" scenes wrong even though such scenes are in the training data?

The moment a scene is converted to feature A (area change, center shift), a lighting scene becomes a point where "the area grows a lot and the center stays put," landing on the same side as approaching scenes, even farther out. A single straight line cannot separate it from approaching scenes. Adding data does not fix it, because the features have lost the information.

Q3. Which human sense goes straight to the cerebral cortex without passing through the thalamus?

Smell. It goes olfactory epithelium → olfactory bulb → piriform cortex. Vision, hearing, and touch pass through the thalamus.

Q4. What do the fruit fly's LPLC2 and feature B in Section 3 have in common?

Both distinguish "an object expanding in all directions" from "simply getting darker." LPLC2 does not respond to dimming without motion. The difference is that feature B was designed by a person, while LPLC2 is a circuit formed during development.

Q5. What is wrong with the judgment "It scored 100% accuracy on the training data, so it will work well in the field too"?

It can fail when it meets situations not in the training data (the lighting changes in Section 4). You must measure it on unseen test data and check whether the situations that can occur in the field are included in the training data.

Colab assignments

Basic

Add "approaching at an angle" (the shadow grows while its center also moves a lot) to the scene types in Code 1. Summarize in a table how the test accuracy of features A and B and of the neural network changes, and explain why in three sentences.

Challenge

Pick one dataset from your own work and design 3 "features people can define." Following the approach of Code 2, compare features + logistic regression against raw data + neural network across numbers of training examples, and write a conclusion, with reasons, on which one to use.

Adjusting difficulty and further learning

BeginnersSkipping the Section 2 table and Section 5 does not break the flow. Just make sure to clearly show the single "Lights dimming" case in Sections 3 & 4
AdvancedDiscuss why, in the hidden neuron images of Section 4, the "before scene" and "after scene" weights take opposite colors (after − before = a filter that looks at change). Continue on to backpropagation in 3-2
Further learningSangsun Ahn, “Data Science for Business” lecture notes (data science in general) · STEP [K-Digital] AI Natural Language Processing (linked to 3-5 and 3-8) · STEP [K-Digital] Fundamentals of Fintech Data Analysis (linked to 3-6)
Materials for instructorsGrading criteria for the check questions and sample answers for the assignments are on the answer key for educators (3-1)

Well done!

You've read all of "3-1 How AI Works." The lab members each have a few words to sum it up for you.

the Professor
ProfessorWe walked the four paths AI works along, side by side.
  • If people choose the features, it's machine learning; if even the features are learned from data, it's deep learning
  • When the lights dimmed, feature A got 0% right and feature B got 100%
  • Read with an eye for what is known versus what is estimated
the Assistant
AssistantSo the key was who decides the rules.
the Fly Researcher
Fly ResearcherI was on the deep learning side from the start, you know.