Dari mimpi sci-fi sampai revolutionize dunia modern
Alan Turing mempublikasikan paper "Computing Machinery and Intelligence" yang mengajukan pertanyaan fundamental: "Can machines think?"
Turing Test menjadi benchmark pertama untuk mengukur kecerdasan mesin. Jika mesin bisa membuat manusia percaya bahwa mereka sedang berbicara dengan manusia lain, maka mesin tersebut dianggap "intelligent".
Dartmouth Conference - John McCarthy, Marvin Minsky, Claude Shannon, dan Nathan Rochester mengorganisir workshop pertama tentang "Artificial Intelligence".
Istilah "Artificial Intelligence" secara resmi lahir di sini. Mereka percaya bahwa "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."
Frank Rosenblatt menciptakan Perceptron, neural network pertama yang bisa belajar dari data!
Perceptron adalah algoritma supervised learning untuk binary classification. Ini adalah building block fundamental dari neural networks modern.
// Perceptron Algorithm (Simplified)
y = activation(w₁x₁ + w₂x₂ + ... + wₙxₙ + bias)
// Update rule
w_new = w_old + learning_rate × (target - output) × input
Minsky & Papert mempublikasikan "Perceptrons" yang menunjukkan limitation dari single-layer perceptron (tidak bisa solve XOR problem).
Funding untuk AI research mengering. Banyak researcher meninggalkan field ini. Periode ini disebut "AI Winter" pertama.
Rumelhart, Hinton, Williams mempopulerkan backpropagation algorithm untuk training multi-layer neural networks.
Backpropagation memungkinkan neural networks dengan multiple hidden layers untuk belajar complex patterns. Ini adalah breakthrough yang membuka jalan untuk deep learning!
// Backpropagation Core Idea
1. Forward Pass: Calculate output
2. Calculate Loss: How far from target?
3. Backward Pass: Calculate gradients (chain rule!)
4. Update Weights: w = w - learning_rate × gradient
5. Repeat until convergence
IBM Deep Blue mengalahkan World Chess Champion Garry Kasparov dalam 6-game match.
Ini adalah pertama kalinya komputer mengalahkan world champion di complex game seperti chess. Deep Blue bisa evaluate 200 million positions per second!
Geoffrey Hinton memperkenalkan Deep Belief Networks (DBNs) dan greedy layer-wise pretraining.
Hinton menunjukkan bahwa deep neural networks bisa di-train effectively dengan unsupervised pretraining followed by supervised fine-tuning. Ini memulai "Deep Learning Revolution"!
Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton memenangkan ImageNet competition dengan margin yang HUGE menggunakan deep CNN.
AlexNet mencapai top-5 error rate 15.3%, jauh lebih baik dari runner-up (26.2%). Ini adalah moment dimana deep learning EXPLODED!
// AlexNet Architecture
Input (224x224x3)
→ Conv1 (96 filters, 11x11) + ReLU + MaxPool
→ Conv2 (256 filters, 5x5) + ReLU + MaxPool
→ Conv3 (384 filters, 3x3) + ReLU
→ Conv4 (384 filters, 3x3) + ReLU
→ Conv5 (256 filters, 3x3) + ReLU + MaxPool
→ FC6 (4096) + ReLU + Dropout
→ FC7 (4096) + ReLU + Dropout
→ FC8 (1000) + Softmax
Ian Goodfellow memperkenalkan GANs, revolutionary approach untuk generative modeling.
GANs terdiri dari dua networks yang "compete": Generator (creates fake data) vs Discriminator (detects fake data). Kompetisi ini menghasilkan incredibly realistic generated images!
DeepMind's AlphaGo mengalahkan world champion Lee Sedol 4-1 di game Go.
Go memiliki 10^170 possible board configurations (lebih banyak dari atoms di universe!). AlphaGo menggunakan kombinasi deep neural networks + Monte Carlo tree search + reinforcement learning.
Vaswani et al. (Google Brain) memperkenalkan Transformer architecture.
Transformers menggunakan self-attention mechanism tanpa recurrence. Ini REVOLUTIONIZED NLP dan menjadi foundation untuk BERT, GPT, dan semua large language models modern!
// Self-Attention Formula
Attention(Q, K, V) = softmax(QK^T / √d_k) × V
Where:
Q = Query (what I'm looking for)
K = Key (what I have to offer)
V = Value (what I'll actually give)
d_k = dimension (for scaling)
Google AI merilis BERT (Bidirectional Encoder Representations from Transformers).
BERT di-pretrain dengan "masked language modeling" - predict missing words in sentence. Ini menghasilkan powerful contextualized word embeddings yang CRUSHING benchmarks!
OpenAI merilis GPT-3, largest language model at the time dengan 175B parameters.
GPT-3 menunjukkan "emergent abilities" - capabilities yang muncul dari scale: few-shot learning, reasoning, code generation, creative writing, dan bahkan simple arithmetic!
Generative AI for images goes MAINSTREAM! Text-to-image models yang bisa generate photorealistic images dari text prompts.
Stable Diffusion: Open-source diffusion model
DALL-E 2: OpenAI's image generator
Midjourney: Artistic AI generation
OpenAI merilis ChatGPT based on GPT-3.5, dengan RLHF (Reinforcement Learning from Human Feedback).
ChatGPT mencapai 1 MILLION users dalam 5 hari (fastest growing app in history!). Ini membawa AI ke mainstream consciousness.
Multimodal AI era begins!
GPT-4: Multimodal (text + images), 1.76 trillion parameters (rumored), human-level performance di many benchmarks.
Claude 3: Anthropic's constitutional AI dengan longer context windows (200k tokens!).
The Current Frontier:
Traditional Programming: Humans write explicit rules
Input + Rules → Output
Machine Learning: Machines learn patterns from data
Input + Output → Model discovers Rules
Supervised Learning: Learn from labeled data
Examples: Classification, Regression
Use cases: Email spam detection, house price prediction, image classification
Unsupervised Learning: Find patterns in unlabeled data
Examples: Clustering, Dimensionality Reduction
Use cases: Customer segmentation, anomaly detection, recommendation systems
Think of neural networks like a team of specialists:
Training = Learning from mistakes!
Network makes prediction → Compare with truth → Adjust weights → Repeat!
Deep Learning adalah neural networks dengan MANY hidden layers (hence "deep").
Why depth matters?
Each layer builds on previous layer → Hierarchical feature learning!
Discriminative Models: "Is this a cat or dog?"
Learn decision boundaries, classify existing data
Generative Models: "Create a new cat image!"
Learn data distribution, generate NEW data
Three Key Factors:
Moore's Law + Data = AI Renaissance!