AILT1001 L2 · Animated Student Slides

L1 Exit Ticket Results

Recap: Course Essentials

Last Week: The Path to Modern AI

Brain ≈86 Billion Biological
Neurons

1943

Artificial Neuron Numbers In → Number Out

1958

Perceptron Learns Weights from Examples

1986

Backpropagation Error Guides Weight Updates

2012

Deep Learning More Layers + Data + GPUs

2017

Transformer Attention Connects Context

2020s

Today

Modern LLMs Transformers +
Pretraining + Scale

Inspired by Biology. Built from Math, Data, and Computing.

Why Does It Matter
How AI Learns and Makes Predictions?

Pilots Do Not Need to Know How to Build Planes

A pilot operating and monitoring the systems in a modern aircraft cockpit

Understanding how planes fly helps pilots recognise CAPABILITIES · LIMITS · WARNING SIGNS

Understanding how AI learns and predicts reveals its CAPABILITIES · LIMITS · POSSIBLE FAILURES and helps you decide when to trust it.

What Powers These Products?

ChatGPT

Claude

DeepSeek

TikTok

Netflix

Spotify

Snapchat Filters

Phone Camera

NEURAL NETWORKS power many features in these products

TODAY: 24 Nov 2026

A day before the Final Test.

You are worried about your test tomorrow

Someone Gives You This Device

You don’t know how it works.
Would you trust it?

Neural Network Inside!

1INPUT NODESreceive the input values

2SYNAPSEScarry signals between nodes

3WEIGHTSstrengthen or weaken each signal

4ARTIFICIAL NEURONScombine inputs and produce one output

Inference: Use the Network

1INPUT NODESreceive the input values

2SYNAPSEScarry signals between nodes

3WEIGHTSstrengthen or weaken each signal

4ARTIFICIAL NEURONScombine inputs and produce one output

So Far, the Knob Positions Were Already SetTraining Determines the Knob Positions

KNOWN TRAINING DATA

AApple Cheung3h sleep · 5h studySCORE75

RRainbow Wong5h sleep · 1h studySCORE82

SSunny Choi10h sleep · 2h studySCORE93

Training: Search for Better Knob Positions

KNOWN TRAINING DATA

AApple Cheung3h sleep · 5h studySCORE75

RRainbow Wong5h sleep · 1h studySCORE82

SSunny Choi10h sleep · 2h studySCORE93

How Can We Search?

TRY100

TRY999,999,901999,999,902999,999,903999,999,904999,999,905999,999,906999,999,907999,999,908999,999,909999,999,910999,999,911999,999,912999,999,913999,999,914999,999,915999,999,916999,999,917999,999,918999,999,919999,999,920999,999,921999,999,922999,999,923999,999,924999,999,925999,999,926999,999,927999,999,928999,999,929999,999,930999,999,931999,999,932999,999,933999,999,934999,999,935999,999,936999,999,937999,999,938999,999,939999,999,940999,999,941999,999,942999,999,943999,999,944999,999,945999,999,946999,999,947999,999,948999,999,949999,999,950999,999,951999,999,952999,999,953999,999,954999,999,955999,999,956999,999,957999,999,958999,999,959999,999,960999,999,961999,999,962999,999,963999,999,964999,999,965999,999,966999,999,967999,999,968999,999,969999,999,970999,999,971999,999,972999,999,973999,999,974999,999,975999,999,976999,999,977999,999,978999,999,979999,999,980999,999,981999,999,982999,999,983999,999,984999,999,985999,999,986999,999,987999,999,988999,999,989999,999,990999,999,991999,999,992999,999,993999,999,994999,999,995999,999,996999,999,997999,999,998999,999,9991,000,000,000

TRY1,000,000,000

10 TRIES / SECOND→≈ 3 YEARS · 2 MONTHS · 2 DAYS

OUR NETWORK9KNOBS

VS

DEEPSEEK-V4 PRO1,700,000,000,000KNOBS

How Did They Train It?

BACKPROPAGATIONprovides a useful direction for updating all the knobs AT THE SAME TIME

Training in Action

Layers of a Neural Network

1INPUT LAYER2 input nodes

2HIDDEN LAYER(S)1 hidden layer · 3 neurons

3OUTPUT LAYER1 output neuron

2+ hidden layers → deep network

Changing the Architecture

1INPUT LAYER2 → 3 input nodes

2HIDDEN LAYER(S)1 → 2 hidden layers

2HIDDEN LAYER(S)3 → 5 neurons / layer

3OUTPUT LAYER1 → 3 output neurons

Different Inputs, Different Outputs

NUMBERS IN

SLEEP8 h

STUDY3 h

NUMBER OUT74/ 100PREDICTED SCORE

OUR EXAMPLEnumbers → predicted score

LOW-RES IMAGE IN A visibly pixelated low-resolution cockpit image

ENHANCED IMAGE OUT

A clearer high-resolution version of the cockpit image

SUPER-RESOLUTIONlow-resolution image → high-resolution image

WORDS INThe deadline was

NEXT WORD OUTyesterday

TEXT PREDICTIONprevious words → next word

Activity

1Ed Login on your phone ↓ 2Camera App ↓ 3Scan QR ↓ 4Submit

ACTIVITY · EXIT TICKET

How Neural Networks Learn