After this lesson you can explain the GAN architecture and the adversarial training dynamic; describe why GANs were difficult to train and how Diffusion Models provide a more stable alternative; distinguish between autoregressive LLMs (GPT family) and diffusion-based image generators; explain what multimodal generative AI means and give examples; and identify both the opportunities and the specific security risks that Generative AI introduces in maritime contexts.
This lesson closes PART 1 of the AI Learning Roadmap by completing the arc begun in Lesson 1: from rule-based systems that follow human instructions, through ML systems that recognise patterns in data, to Generative AI systems that create new content — text, images, code, audio, video — that did not exist before.
The story begins in 2014 with a paper published at NIPS by Ian Goodfellow and colleagues that introduced the Generative Adversarial Network — an architecture so novel that Yann LeCun, Chief AI Scientist at Meta, called it "the most interesting idea in the last 10 years in machine learning." The story continues through Diffusion Models, which became the dominant image generation technique by 2022, and culminates in the multimodal large language models that define AI in 2025–26.
For maritime professionals, the practical implications run in two directions: Generative AI creates new tools for maritime operations (document drafting, simulation, training), and it creates new threats (deepfake communications, AI-generated phishing, synthetic sensor data). Both directions require understanding how these systems work.
1. The GAN Breakthrough — Goodfellow et al., 2014
Before GANs, generating realistic images with neural networks was an unsolved problem. Variational Autoencoders (VAEs, Kingma and Welling, 2013) could generate images but produced blurry outputs. Autoregressive pixel-by-pixel generation was computationally expensive and produced inconsistent results. The field did not have a working architecture for high-quality image synthesis.
Ian Goodfellow and his co-authors (Pouget-Abadie, Mirza, Xu, Warde-Farley, Ozair, Courville, Bengio) introduced the Generative Adversarial Network (GAN) in their paper "Generative Adversarial Nets," presented at NIPS 2014 (arXiv:1406.2661). The paper describes a framework where two networks compete in a minimax game — an idea Goodfellow reportedly conceived while arguing about generative models at a bar in Montreal.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). "Generative Adversarial Nets." Advances in Neural Information Processing Systems (NIPS) 2014. arXiv:1406.2661. The paper introduced a new training framework for generative models and proved that the adversarial game has a unique global optimum where the generator perfectly replicates the data distribution and the discriminator can no longer distinguish real from generated samples.
2. How GANs Work — Generator vs. Discriminator
A GAN consists of two neural networks trained simultaneously in an adversarial game:
🎨 Generator (G)
Takes a random noise vector z (sampled from a Gaussian distribution) as input and produces a synthetic sample — an image, a data point, a sequence — that it tries to make indistinguishable from real data.
The Generator never sees real data directly — it only receives the Discriminator's feedback signal about whether its outputs are convincing.
🔍 Discriminator (D)
Takes a sample — either real (from the training dataset) or fake (generated by G) — and outputs a probability that the sample is real. It is a binary classifier.
The Discriminator is trained to correctly distinguish real from fake. As G improves, D must become more discerning. As D improves, G must produce more convincing fakes.
The training dynamic: G and D are trained alternately. When training G, D's parameters are frozen — G is updated to maximise D's error rate (fool D). When training D, G's parameters are frozen — D is updated to correctly classify real vs. fake. The loss function for G is: minimise the probability that D correctly identifies G's outputs as fake. Formally, this is a minimax game: min_G max_D V(D,G).
In the ideal equilibrium (Nash equilibrium), G has learned the true data distribution exactly and D can do no better than random guessing (50% accuracy on real vs. fake). In practice, this equilibrium is hard to reach: GANs are notoriously difficult to train, often suffering from mode collapse (G learns to produce only a few types of samples that fool D) and training instability (the discriminator overwhelms the generator early in training).
3. GAN Variants and Limitations
The original GAN generated 28×28 pixel digit images. Subsequent variants scaled to high-resolution photorealistic images:
Despite these advances, GANs remained difficult to train reliably. Mode collapse, training instability, and the need for careful hyperparameter tuning made them impractical for many applications. By 2022, a fundamentally different generative approach had surpassed GANs for most image generation tasks: Diffusion Models.
4. Diffusion Models — A More Stable Path to Generation
Diffusion Models (Ho, Jain, and Abbeel, "Denoising Diffusion Probabilistic Models," NeurIPS 2020, arXiv:2006.11239) take a different mathematical approach to generation. Instead of a generator-discriminator adversarial game, a diffusion model learns to reverse a gradual noising process.
Forward process (training): Take a real image and gradually add Gaussian noise over T steps (typically 1,000 steps) until the image becomes pure noise. This process is fixed and requires no learning.
Reverse process (generation): Train a neural network (typically a U-Net) to predict the noise added at each step, given the current noisy image and the step number. After training, generation is performed by starting from pure random noise and running the learned denoising network T times — gradually removing noise until a clean image emerges.
Diffusion models train more stably than GANs (no adversarial game, no mode collapse), generate higher-quality and more diverse outputs, and have become the dominant paradigm for image and video generation as of 2022–24. The trade-off is inference speed: generating a single image requires 50–1,000 denoising steps, making diffusion models slower at inference than GANs (though techniques like DDIM reduce this significantly).
5. Large Language Models — Text as the Universal Interface
Parallel to the image generation story, the Transformer architecture enabled a new approach to text generation through autoregressive language modelling: train a Transformer to predict the next token in a sequence, given all previous tokens. At scale — GPT-3's 175 billion parameters trained on 300 billion tokens from the internet — this simple objective produces a model that can write code, translate languages, summarise documents, answer questions, and reason through multi-step problems.
The LLM text generation process at inference: given a prompt (a sequence of input tokens), the model predicts the next token by sampling from its learned probability distribution. The sampled token is appended to the prompt, and the process repeats — autoregressively generating the response one token at a time. The temperature parameter controls the randomness of sampling: temperature = 0 gives the always most-likely token (deterministic, repetitive); temperature = 1 gives samples proportional to the model's learned distribution (creative, sometimes wrong).
LLM의 텍스트 생성 능력이 어떻게 Agentic AI로 확장되는지를 5편 시리즈로 다룹니다. 이 레슨의 섹션 5와 6을 이해했다면 Agent AI 시리즈로 자연스럽게 이어집니다.
LLM 기반 생성형 에이전트가 25개의 가상 인물을 시뮬레이션하며 인간과 유사한 행동을 보이는 실험 연구입니다. Generative AI의 에이전트 응용을 탐구하는 중요한 논문입니다.
6. Multimodal Generative AI — Text, Image, Audio, and Beyond
The defining characteristic of the 2024–26 generation of AI systems is multimodality: the ability to accept and generate combinations of text, images, audio, and structured data within a single model. This is enabled by the same Transformer architecture applied to tokenised representations of multiple modalities simultaneously.
멀티모달 생성형 AI가 기업 현장에서 어떻게 배포되고, 조직적으로 어떤 준비가 필요한지를 2027년 관점에서 다룹니다. 이번 레슨의 섹션 6에서 소개된 시스템들의 실무 적용 사례를 확인하세요.
7. Maritime Connection — Generative AI as Tool and Threat
Generative AI creates a dual-use challenge for maritime organisations that is qualitatively different from earlier AI generations. The same capability that drafts a port state control response also generates convincing spear-phishing emails. The same model that generates vessel training simulations also generates synthetic AIS tracks for route deception.
The NIST AI RMF GOVERN and MANAGE functions address exactly this challenge: establishing policies for permitted and prohibited AI uses in organisational context, and monitoring deployed AI systems for misuse. For maritime operators, this means having an AI Use Policy that specifies which Generative AI tools are approved for which tasks, what data can be submitted to external AI systems, and how AI-generated content should be reviewed before operational use.
During E26 CRSI 프로젝트 수행, I observe that maritime organisations are adopting Generative AI tools at the employee level faster than IT/security departments are establishing governance frameworks for their use. The typical scenario: individual employees use ChatGPT or Claude to draft ISM-related documents, translate communications, or summarise regulatory texts — without the organisation having formally assessed what data is being submitted to these systems.
The immediate cybersecurity risk is data exfiltration through AI prompts: if a crew member pastes the vessel's OT network topology, maintenance records, or incident reports into a public LLM to get a summary, that data leaves the organisation's control. Most Generative AI providers' default data retention policies allow the provider to use submitted content for model improvement — which may conflict with the confidentiality requirements of Class society assessments and IACS UR E26 compliance documentation.
The practical recommendation: establish an AI Use Policy that clearly delineates (1) which AI tools are approved (enterprise versions with no data retention, or on-premise deployments), (2) what categories of data may not be submitted to any AI system without explicit approval, and (3) how AI-generated content is reviewed before operational or regulatory use. This policy should be integrated with the vessel's cybersecurity management system under IACS UR E26 Section 5.1 documentation requirements.
- The GAN (Goodfellow et al., 2014) introduced the adversarial training framework: Generator and Discriminator trained in a minimax game. First architecture to produce photorealistic image synthesis.
- GANs suffer from mode collapse and training instability. Diffusion Models (Ho et al., 2020) provide a more stable alternative using iterative denoising of Gaussian noise.
- Stable Diffusion democratised image generation by running diffusion in latent space, enabling consumer GPU deployment. Open model weights mean the capability is widely accessible — including to attackers.
- LLMs generate text autoregressively: one token at a time, sampled from the model's learned probability distribution. Temperature controls the randomness of sampling.
- Multimodal GenAI (GPT-4V, Gemini, Claude) accepts and generates across text, images, audio, and video — enabling maritime applications from document analysis to vessel training simulation.
- Generative AI is dual-use: the same capabilities that assist maritime operations also enable AI-enhanced phishing, synthetic AIS manipulation, deepfake CEO fraud, and LLM-assisted exploit development.
- Governance (NIST AI RMF GOVERN function, IACS UR E26 Section 5.1) requires an explicit AI Use Policy covering approved tools, restricted data categories, and content review protocols.
AI Fundamentals — All 5 Lessons Done
You have now traced the arc from Rule-Based AI through Machine Learning, Deep Learning (CNN/RNN), Computer Vision, and Generative AI. You have the conceptual foundation to understand every AI system you will encounter in the remainder of this roadmap.
View Full Roadmap → PART 2PART 2 — Transformer Architecture & Language Models opens the Transformer black box introduced in Lesson 1: attention heads, positional encoding, the encoder-decoder structure, scaling laws, and why the architecture generalises so far beyond its original machine translation application. If Lesson 1 was the map of the territory, PART 2 is the expedition into it.
PART 2 also introduces the NIST AI RMF MAP function in the context of LLM risk categorisation — a practical framework for assessing and documenting the risk profile of AI systems in maritime operational and cybersecurity contexts.
- Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al. (2014). "Generative Adversarial Nets." NIPS 2014. arXiv:1406.2661.
- Radford, A., Metz, L., and Chintala, S. (2015). "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks." arXiv:1511.06434 (DCGAN).
- Karras, T., Laine, S., and Aila, T. (2019). "A Style-Based Generator Architecture for Generative Adversarial Networks." CVPR 2019. arXiv:1812.04948 (StyleGAN).
- Ho, J., Jain, A., and Abbeel, P. (2020). "Denoising Diffusion Probabilistic Models." NeurIPS 2020. arXiv:2006.11239.
- Rombach, R., Blattmann, A., Lorenz, D., et al. (2022). "High-Resolution Image Synthesis with Latent Diffusion Models." CVPR 2022. arXiv:2112.10752 (Stable Diffusion).
- Radford, A., Kim, J., Hallacy, C., et al. (2021). "Learning Transferable Visual Models From Natural Language Supervision." arXiv:2103.00020 (CLIP).
- OpenAI (2024). GPT-4 Technical Report. arXiv:2303.08774.
- NIST AI RMF 1.0 (January 2023). airc.nist.gov/RMF — GOVERN and MANAGE functions: AI Use Policy and operational oversight.
- IACS UR E26 (Rev.3, 2024) — Section 5.1: Documentation requirements for cybersecurity management systems. iacs.org.uk
⚓ Join the ShipPaulJobs Community
Join →
Comments
Post a Comment