# Sieci i modele generatywne

EN: **Neural architectures and generative models** · id `neural`

Path: [AI Atlas](https://aiatlasgraph.com/llms.txt) › Sieci i modele generatywne

## Description

- PL: Rodziny modeli i ich mechanizmy. Nie każdy model generatywny jest modelem językowym.
- EN: Model families and their mechanisms. Not every generative model is a language model.

## In one sentence

- PL: To sposoby budowania modeli, które uczą się wzorów w danych, przewidują wyniki i tworzą nowe treści.
- EN: These are ways to build models that learn patterns in data, predict outcomes and create new content.

## Problem it solves

- PL: Nie da się ręcznie rozpisać wszystkich reguł potrzebnych do rozpoznawania zdjęć, rozumienia tekstu lub tworzenia obrazów.
- EN: You cannot manually write every rule needed to recognise photos, process text or create images.

## What it does

- PL: Połączone części modelu przekształcają dane, a uczenie dostosowuje ich działanie; różne układy odpowiadają różnym zadaniom i kosztom.
- EN: Connected model components transform data, while training adjusts their behaviour; different designs suit different tasks and costs.

## Everyday example

- PL: Jedna aplikacja rozpoznaje przedmioty na zdjęciu, a inna tworzy ilustrację z opisu, korzystając z inaczej zbudowanych modeli.
- EN: One app recognises objects in a photo while another creates an illustration from a description, using differently structured models.

## This concerns you if you…

- PL: chcesz zrozumieć, jak zbudowany jest model
- PL: porównujesz modele rozpoznające dane i tworzące treści
- EN: want to understand how a model is built
- EN: compare models that recognise data and generate content

## Keywords

MLP · backpropagation · CNN · RNN · Transformers · attention mechanisms · graph neural networks · representation learning · diffusion models · VAE · GAN · Mixture of Experts · test-time compute

## How it works

- PL: Parametry sieci dopasowuje się przez minimalizację funkcji celu, budując reprezentacje w kolejnych warstwach. Architektury przestrzenne, sekwencyjne i generatywne różnią się sposobem wymiany informacji oraz procedurą uzyskiwania przewidywań lub próbek.
- EN: Network parameters are fitted by minimising an objective, building representations through successive layers. Spatial, sequential and generative architectures differ in information exchange and procedures for obtaining predictions or samples.

## Limits

- PL: Złożoność modelu zwiększa wymagania dotyczące danych, pamięci i obliczeń. Niski błąd treningowy nie gwarantuje uogólniania; generowana próbka nie jest dowodem poprawności jej treści.
- EN: Model complexity increases data, memory and computation requirements. Low training error does not guarantee generalisation; generated samples do not establish content correctness.

## Key terms

- PL: **parametr uczony**: Wartość dostosowywana podczas treningu do realizacji funkcji celu.
- PL: **architektura**: Układ operacji i połączeń określający sposób przetwarzania danych.
- EN: **learned parameter**: A value adjusted during training to fulfil an objective.
- EN: **architecture**: The arrangement of operations and connections defining data processing.

## Know first

- [Algebra liniowa: wektory i macierze (Linear algebra: vectors and matrices)](https://aiatlasgraph.com/t/linearalgebra.md)
- [Spadek gradientu i optymalizacja ciągła (Gradient descent and continuous optimisation)](https://aiatlasgraph.com/t/mathoptimization.md)

## Subtopics (8 direct, 99 in total)

- [Architektury efektywne (Efficient architectures)](https://aiatlasgraph.com/t/efficientarch.md): Designing a model around cost and scale.
  - [Architektury hybrydowe: uwaga + warstwy liniowe (Hybrid architectures: attention plus linear/recurrent layers)](https://aiatlasgraph.com/t/hybridattention.md): Models combining full-attention layers with recurrent, state-space or linear-attention layers. Fixed-size-state layers can reduce memory cost on long sequences, while periodic full-attention layers preserve direct interactions between distant tokens; exact ratios depend on the architecture.
  - [Mixture of Experts (MoE) — mieszanka ekspertów (Mixture of Experts (MoE))](https://aiatlasgraph.com/t/moe.md): Layers containing many expert subnetworks, where a router sends each token to only a selected subset. Sparse parameter activation can reduce compute per token compared with a dense model of similar total parameter count, while weight storage, communication and load balancing remain important costs.
  - [Modele przestrzeni stanów (Mamba) (State-space models (Mamba))](https://aiatlasgraph.com/t/ssm.md): A family of sequence models that update a fixed-size state instead of storing a full attention matrix or KV cache. Mamba introduces a selective, input-dependent state mechanism; its authors report linear scaling with sequence length and competitive results across several modalities.
  - [Uwaga lokalna z przesuwnym oknem (Sliding-window (local) attention)](https://aiatlasgraph.com/t/slidingwindow.md): Local attention in which a token interacts only with a bounded window of neighbouring tokens rather than the entire sequence. With a fixed window, total compute grows linearly with sequence length, and the KV memory used for generation can be bounded by the window size. Some architectures combine it with global or full attention.
- [Podstawy sieci neuronowych (Neural network fundamentals)](https://aiatlasgraph.com/t/neuralbasics.md): Layers, loss functions and representation learning.
  - [Funkcje aktywacji (ReLU, GELU, SwiGLU) (Activation functions (ReLU, GELU, SwiGLU))](https://aiatlasgraph.com/t/activations.md): Non-linear functions applied to neuron outputs — without them, a composition of linear layers remains a linear transformation. ReLU is common in vision networks; transformers often use GELU and gated variants such as SwiGLU in feed-forward blocks.
  - [Funkcje straty (entropia krzyżowa, MSE) (Loss functions (cross-entropy, MSE))](https://aiatlasgraph.com/t/lossfunctions.md): The number training minimises: how far predictions are from the target. Cross-entropy is used for classification and next-token prediction; mean squared error (MSE) for regression and, among others, noise prediction in diffusion models.
  - [Normalizacja warstw (BatchNorm, LayerNorm, RMSNorm) (Normalisation layers (BatchNorm, LayerNorm, RMSNorm))](https://aiatlasgraph.com/t/normalization.md): Layers that rescale activations and often help stabilise deep-network training. BatchNorm uses mini-batch statistics and is typical in CNNs; LayerNorm normalises features within one example and is common in transformers; RMSNorm simplifies LayerNorm and is often used in newer LLMs, usually in a pre-norm layout.
  - [Perceptron wielowarstwowy (MLP) (Multilayer perceptron (MLP))](https://aiatlasgraph.com/t/mlp.md): A network of fully connected layers that alternate linear transformations with non-linear activation functions. The MLP is a fundamental neural model and a standard component of feed-forward blocks in transformer architectures.
  - [Połączenia resztkowe (skip connections) (Residual (skip) connections)](https://aiatlasgraph.com/t/residual.md): Shortcut connections that add a block's input to its output, allowing the layer to learn a correction relative to the input representation. They enabled effective training of very deep ResNets and are a common component of many modern deep architectures, including transformers.
  - [Propagacja wsteczna (backpropagation) (Backpropagation)](https://aiatlasgraph.com/t/backprop.md): An algorithm for computing the gradient of the loss with respect to network parameters: the forward pass computes the output, while the backward pass propagates derivatives through the computation graph using the chain rule. Gradient-based optimisers use these gradients; libraries usually obtain them through automatic differentiation.
  - [Regularyzacja sieci: dropout i weight decay (Neural network regularisation: dropout and weight decay)](https://aiatlasgraph.com/t/nnregularization.md): Techniques that curb overfitting in deep networks. Dropout randomly switches off units during training; weight decay penalises large weights; early stopping and data augmentation belong to the same family.
- [Reprezentacje i embeddingi (Representation learning)](https://aiatlasgraph.com/t/representations.md): Encoding meaning or features as vectors.
  - [Autoenkodery (ściskanie i odtwarzanie danych) (Autoencoders)](https://aiatlasgraph.com/t/autoencoders.md): An encoder–decoder network trained to reconstruct its input through a narrow hidden layer; a nonlinear PCA and the basis of VAEs and sparse autoencoders.
  - [Embeddingi zdań i dokumentów (modele embeddingowe) (Sentence and document embeddings (embedding models))](https://aiatlasgraph.com/t/textembeddings.md): Models that turn a whole sentence or document chunk into a vector compared, for example, with cosine similarity. They are commonly used in semantic retrieval and RAG pipelines, as well as clustering and deduplication; embedding models are compared on benchmarks such as MTEB. Cross-encoders for reranking and multi-vector representations complement them.
  - [Osadzenia słów (word2vec, GloVe) (Word embeddings (word2vec, GloVe))](https://aiatlasgraph.com/t/wordembeddings.md): Static vectors assigned to words and learned from regularities in their contexts, so that similar uses receive nearby representations. word2vec and GloVe were important steps in the development of text representations; a single vector per word does not distinguish meanings that depend on the particular context.
  - [Uczenie metryk podobieństwa (Deep metric learning)](https://aiatlasgraph.com/t/metriclearning.md): Learning an embedding space where distance reflects semantic similarity, using pairwise, triplet or proxy-based losses.
- [Rodziny modeli generatywnych (Generative model families)](https://aiatlasgraph.com/t/generative.md): Different ways of learning a distribution and sampling from it.
  - [Autoenkodery wariacyjne (VAE) (Variational autoencoders (VAE))](https://aiatlasgraph.com/t/vae.md): A model with an encoder that compresses data into a latent distribution and a decoder that reconstructs it from samples; trained by maximising the evidence lower bound (ELBO) with the reparameterisation trick. Today it often works behind the scenes — as the image compressor in latent diffusion.
    - [Dyskretne kody obrazu i dźwięku (VQ-VAE, VQGAN) (Vector-quantised autoencoders (VQ-VAE, VQGAN))](https://aiatlasgraph.com/t/vqvae.md): An autoencoder whose encoder outputs indices of the nearest vectors in a learned codebook, producing discrete tokens for generative models.
    - [Rozplątane reprezentacje (β-VAE) (Disentangled representations (β-VAE))](https://aiatlasgraph.com/t/betavae.md): A VAE variant with weight β>1 on the KL term, strengthening independence of latent variables at the cost of reconstruction quality.
  - [Autoregresyjne generowanie obrazu i dźwięku (Autoregressive image and audio generation)](https://aiatlasgraph.com/t/argen.md): Models generating data as a sequence of elements (pixels, samples, tokens or scales), each conditioned on earlier ones; a rival and complement to diffusion.
    - [Generowanie piksel po pikselu i próbka po próbce (PixelCNN, WaveNet) (Pixel- and sample-level autoregressive models (PixelCNN, WaveNet))](https://aiatlasgraph.com/t/pixelwavenet.md): Autoregressive convolutional networks modelling pixels or audio samples sequentially using masked or dilated causal convolutions.
    - [Maskowane transformery generatywne (MaskGIT) (Masked generative image transformers (MaskGIT))](https://aiatlasgraph.com/t/maskgit.md): A bidirectional transformer generating image tokens in parallel over a few iterations, unmasking the most confident predictions from a masked grid.
    - [Przewidywanie kolejnej skali (od szkicu do detalu) (Next-scale prediction (visual autoregressive modelling, VAR))](https://aiatlasgraph.com/t/var.md): Autoregression over scales: a transformer predicts a whole higher-resolution token map conditioned on all lower ones, instead of tokens in raster order.
  - [Generatywne sieci przeciwstawne (GAN) (Generative adversarial networks (GAN))](https://aiatlasgraph.com/t/gan.md): Two networks trained in an adversarial game: a generator creates samples and a discriminator distinguishes them from real data; the discriminator's signal trains the generator to better match the data distribution. Training can be unstable and suffer from mode collapse.
    - [Generator sterowany stylem (StyleGAN) (Style-based GAN generators (StyleGAN))](https://aiatlasgraph.com/t/stylegan.md): A GAN generator architecture where a mapping network produces a style vector injected into every synthesis layer, separating coarse attributes from fine detail.
    - [Przekładanie obrazu na obraz (pix2pix, CycleGAN) (Image-to-image translation (pix2pix, CycleGAN))](https://aiatlasgraph.com/t/imagetranslation.md): Conditional GANs learning mappings between image domains, from paired examples (pix2pix) or without pairs via a cycle-consistency loss (CycleGAN).
    - [Stabilne trenowanie GAN (odległość Wassersteina, kara gradientu) (Stable GAN training: Wasserstein loss and gradient penalty)](https://aiatlasgraph.com/t/wgan.md): A family of GAN objective changes where a critic estimates the Wasserstein distance between distributions, with the Lipschitz constraint enforced by a gradient penalty.
  - [Modele dyfuzyjne (Diffusion models)](https://aiatlasgraph.com/t/diffusion.md): Generating by gradual denoising.
    - [Dyfuzja w przestrzeni pikseli (Pixel-space diffusion)](https://aiatlasgraph.com/t/pixeldiffusion.md): Diffusion trained and sampled directly on the pixel grid, without an autoencoder.
      - [Klepsydrowy transformer dyfuzyjny (Hourglass diffusion transformer)](https://aiatlasgraph.com/t/hourglassdit.md): An hourglass-shaped transformer that scales linearly with pixel count.
      - [Przesunięcie harmonogramu szumu (Noise-schedule shifting for high-resolution pixels)](https://aiatlasgraph.com/t/noiseshift.md): An adjustment of log signal-to-noise ratio as the pixel count grows.
      - [Transformer pikselowy o dwóch poziomach (Dual-level pixel diffusion transformer)](https://aiatlasgraph.com/t/pixeldit.md): A single-stage diffusion transformer with a patch level and a pixel level.
    - [Dyfuzja w przestrzeni ukrytej (Stable Diffusion) (Latent diffusion (Stable Diffusion))](https://aiatlasgraph.com/t/latentdiffusion.md): Diffusion performed in an autoencoder's compressed latent space rather than directly in pixel space, reducing computation; text conditioning can be implemented through cross-attention. This design is used by Latent Diffusion models, including Stable Diffusion.
    - [Dyfuzyjne modele językowe (Diffusion language models)](https://aiatlasgraph.com/t/dllm.md): A family of language models that learn to reverse token masking or noising, rather than only next-token prediction.
      - [Dyfuzja blokowa (Block diffusion language models)](https://aiatlasgraph.com/t/blockdiffusion.md): An autoregressive distribution over blocks, with discrete diffusion inside a block and causal attention across blocks.
      - [Dyfuzja ciągła na zanurzeniach słów (Continuous diffusion on word embeddings)](https://aiatlasgraph.com/t/embdiffusion.md): Continuous diffusion in embedding space: Gaussian noise on vectors, then rounding back to a token.
      - [Dyskretna dyfuzja ze score entropy (Score entropy discrete diffusion)](https://aiatlasgraph.com/t/sedd.md): Discrete diffusion trained with score entropy, which matches ratios of token probabilities.
      - [Dyskretna dyfuzja ze strukturalną macierzą przejść (Structured discrete diffusion (D3PM))](https://aiatlasgraph.com/t/d3pm.md): Discrete diffusion with a transition matrix: uniform, neighbour-based, or with a mask as an absorbing state.
      - [Maskowana dyfuzja tekstu (Masked diffusion language models)](https://aiatlasgraph.com/t/maskdiffusion.md): Diffusion whose only corruption is a mask token: the objective is a weighted mixture of masked-language losses.
        - [Próbkowanie z ponownym maskowaniem (Remasking samplers for masked diffusion)](https://aiatlasgraph.com/t/remasking.md): Masked-diffusion samplers that can mask previously predicted tokens again and revise them over subsequent steps.
    - [Kaskadowe modele dyfuzyjne (Cascaded diffusion models)](https://aiatlasgraph.com/t/cascadediff.md): A chain of diffusion models at increasing resolution, from sketch to detail.
      - [Augmentacja obrazu warunkującego (Conditioning augmentation for cascaded diffusion)](https://aiatlasgraph.com/t/condaugment.md): Corrupting the low-resolution input while training super-resolution models.
    - [Sterowanie promptem bez klasyfikatora (classifier-free guidance) (Classifier-free guidance)](https://aiatlasgraph.com/t/cfg.md): A technique to strengthen adherence to the prompt: one model learns to generate with and without the condition (dropped randomly in training), and at sampling time the output is pushed toward the conditional version. The guidance scale is the familiar “prompt fidelity vs diversity” knob.
    - [Szybkie generowanie w kilku krokach (Fast and few-step diffusion sampling)](https://aiatlasgraph.com/t/fewstepdiffusion.md): Methods for shortening diffusion sampling: better solvers without retraining and distillation into few-step and one-step models.
      - [Modele spójności (generowanie jednym krokiem) (Consistency models)](https://aiatlasgraph.com/t/consistencymodels.md): Models trained to map any point on a denoising trajectory to its endpoint, enabling one- and few-step sampling; trained by distillation or standalone.
      - [Przeciwstawna destylacja dyfuzji (generowanie w 1–4 krokach) (Adversarial diffusion distillation)](https://aiatlasgraph.com/t/advdistill.md): Distilling large diffusion models to 1–4 steps by combining a score-based teacher signal with an adversarial loss.
      - [Stopniowa destylacja kroków (Progressive distillation of diffusion samplers)](https://aiatlasgraph.com/t/progdistill.md): Iterative distillation where a student learns to reproduce two steps of the teacher's deterministic sampler in one, halving the step count each round.
      - [Szybsze samplery bez ponownego treningu (DDIM, solvery ODE) (Fast training-free samplers (DDIM, ODE solvers))](https://aiatlasgraph.com/t/ddim.md): Deterministic or less stochastic sampling procedures for trained diffusion models, based on non-Markovian processes and probability-flow ODE solvers.
    - [Transformery w modelach dyfuzyjnych i przepływowych (DiT) (Transformers for diffusion and flow models (DiT))](https://aiatlasgraph.com/t/dit.md): Generative architectures in which a transformer processes patches of an image representation instead of a conventional U-Net backbone. DiT applied this design to latent diffusion, while newer variants also use transformers in flow-trained models.
    - [Zaszumianie i odszumianie krok po kroku (DDPM) (Forward noising and learned denoising (DDPM))](https://aiatlasgraph.com/t/ddpm.md): The basic construction of denoising diffusion probabilistic models: a forward process gradually adds noise to data, while a model learns to reverse the corruption, often by predicting the added noise. Sampling-step count affects cost and quality, and separate fast-sampling or distillation techniques can reduce it.
  - [Modele oparte na energii (Energy-based models (EBM))](https://aiatlasgraph.com/t/ebm.md): Models defining an unnormalised density through an energy function, trained via MCMC maximum likelihood, score matching or noise-contrastive estimation.
  - [Przepływy normalizujące i flow matching (Normalising flows and flow matching)](https://aiatlasgraph.com/t/flowmatching.md): Models that turn simple noise into data with an invertible transform (normalising flows) or a learned velocity field along a path from noise to data (flow matching, rectified flow). Flow matching includes diffusion as a special case and underpins many new image models, e.g. Stable Diffusion 3.
    - [Prostowanie przepływu (rectified flow) (Rectified flow)](https://aiatlasgraph.com/t/rectifiedflow.md): A flow-matching variant learning a velocity field along straight lines between noise and data, with optional repeated straightening (reflow).
    - [Przepływy odwracalne z warstwami sprzężenia (RealNVP, Glow) (Coupling-layer normalising flows (RealNVP, Glow))](https://aiatlasgraph.com/t/normflows.md): Generators built from invertible transformations with tractable Jacobian determinants, giving exact log-likelihood via the change-of-variables formula.
- [Rozumowanie w modelach językowych (Reasoning in language models)](https://aiatlasgraph.com/t/reasoningmodels.md): Techniques and model families for tasks that require multi-step reasoning: from chain-of-thought prompting and larger inference-time compute budgets to training reasoning models with RL, SFT and distillation.
  - [Efektywne rozumowanie (przeciw „overthinkingowi”) (Efficient reasoning)](https://aiatlasgraph.com/t/efficientreasoning.md): Methods that cut the cost and latency of reasoning: a model should not write thousands of thinking tokens for a simple question.
    - [Adaptacyjne i hybrydowe tryby myślenia (Adaptive and hybrid thinking modes)](https://aiatlasgraph.com/t/adaptivethinking.md): One model can answer fast or think longer and decides by itself (or on request) how much reasoning a question needs.
    - [Skracanie łańcucha myśli (Reasoning length reduction)](https://aiatlasgraph.com/t/concisereasoning.md): Training or prompting aimed at reducing reasoning length while preserving answer quality as much as possible — for example via a length-cost signal, fine-tuning on concise traces, or brevity instructions.
  - [Monitorowanie łańcucha myśli (Chain-of-thought monitoring)](https://aiatlasgraph.com/t/cotmonitoring.md): Analysing a model's written chain of thought (automatically or by a human) to detect signals of undesirable behaviour. Its usefulness depends in part on the monitorability and faithfulness of the reasoning trace, which are not guaranteed.
  - [Obliczenia w czasie wnioskowania (Test-time compute scaling)](https://aiatlasgraph.com/t/testtime.md): Giving a model a bigger compute budget at answer time — without changing weights: many attempts in parallel, successive revisions, verifier-guided search, or simply thinking longer.
    - [Iteracyjne poprawianie odpowiedzi (Sequential self-revision)](https://aiatlasgraph.com/t/sequentialrevision.md): The model reviews and revises its own answer in turns, each version building on the previous — compute grows in depth rather than width.
    - [Przeszukiwanie kroków z weryfikatorem (Verifier-guided search)](https://aiatlasgraph.com/t/verifiersearch.md): The solution is built step by step while a verifier (e.g. a process reward model) scores partial paths and decides which to expand.
    - [Sterowanie budżetem myślenia (Thinking budget control)](https://aiatlasgraph.com/t/budgetcontrol.md): Explicitly setting how many tokens or how much time the model may spend reasoning — via a parameter, a cap, or forcing it to stop or continue thinking.
    - [Wiele prób i głosowanie (Parallel sampling and voting)](https://aiatlasgraph.com/t/parallelsampling.md): The model generates many independent solutions in parallel; the answer is picked by majority vote (self-consistency) or by a verifier score (best-of-N).
  - [Pętle głębokości i myśli ukryte (Recurrent depth and latent thoughts)](https://aiatlasgraph.com/t/latentloops.md): Scaling answer-time compute by repeating a block or a continuous state, not by extra tokens.
    - [Adaptacyjne zatrzymanie obliczeń (Adaptive computation time)](https://aiatlasgraph.com/t/acthalt.md): A differentiable rule that teaches a recurrent network how many steps to take between input and output.
    - [Ciągłe myśli (Continuous chain of thought)](https://aiatlasgraph.com/t/coconut.md): A reasoning chain in which the hidden state returns as an embedding, without decoding to a token.
    - [Głębokość rekurencyjna (Recurrent-depth language models)](https://aiatlasgraph.com/t/recurrentdepth.md): A language model scales answer compute by unrolling a shared recurrent block in hidden space.
    - [Transformery w pętli (Looped transformers)](https://aiatlasgraph.com/t/loopedtf.md): A shared transformer block applied recurrently to sequence representations, sometimes with per-position halting.
  - [Trenowanie modeli rozumujących (Training reasoning models)](https://aiatlasgraph.com/t/reasoningtraining.md): How to teach a model long, self-correcting reasoning: RL rewarded for correct results, SFT on long chains of thought, and distillation from a larger model.
    - [Destylacja rozumowania do mniejszych modeli (Reasoning distillation)](https://aiatlasgraph.com/t/reasoningdistill.md): A large reasoning model generates reasoning traces that a smaller model is fine-tuned on. A cheap way to give a small model reasoning ability.
    - [RL z weryfikowalnymi nagrodami (RLVR) (Reinforcement learning with verifiable rewards)](https://aiatlasgraph.com/t/rlvr.md): The model receives reward from an automatically verifiable signal — such as answer correctness, passing code tests, or satisfying a checkable condition — rather than relying only on a learned preference model. RLVR is an important post-training approach for reasoning models, especially in domains with unambiguous verification.
      - [Rekurencyjne samoulepszanie agentów terminalowych (Recursive self-improvement of terminal agents)](https://aiatlasgraph.com/t/terminalrsi.md): A post-training method alternating supervised learning on verified successful trajectories with PPO using an inherited value model.
    - [SFT na długich łańcuchach myśli (Long chain-of-thought fine-tuning)](https://aiatlasgraph.com/t/longcotsft.md): Supervised fine-tuning on curated examples with long reasoning traces; often a “cold start” before RL, or a cheap route to a reasoning model from few examples.
  - [Weryfikatory rozumowania (Reasoning verifiers)](https://aiatlasgraph.com/t/reasoningverifiers.md): Models or programs that judge whether reasoning is correct. Used both as a training reward and to pick the best answer at inference time.
    - [Ocena każdego kroku (PRM) (Process reward models)](https://aiatlasgraph.com/t/prm.md): A reward model scores each reasoning step separately, not just the final answer. Catches a mistake halfway through a solution.
    - [Sprawdzanie końcowego wyniku (Outcome verification)](https://aiatlasgraph.com/t/outcomeverify.md): Only the final answer is judged: by comparing with a reference, running tests, or an outcome reward model (ORM). Simpler and cheaper than step scoring.
  - [Łańcuch myśli (chain-of-thought) (Chain-of-thought prompting)](https://aiatlasgraph.com/t/cot.md): Getting the model to write out intermediate reasoning steps before the answer. The simplest way to improve multi-step tasks — no training required.
    - [Drzewo myśli (Tree of Thoughts) (Tree of Thoughts)](https://aiatlasgraph.com/t/treeofthoughts.md): The model expands several alternative reasoning paths, evaluates which look promising and can backtrack — tree search instead of one line of thought.
    - [Łańcuch myśli z przykładami (Few-shot chain-of-thought)](https://aiatlasgraph.com/t/fewshotcot.md): The prompt includes a few example problems with step-by-step worked solutions; the model imitates that style on a new problem.
    - [„Pomyśl krok po kroku” bez przykładów (Zero-shot chain-of-thought)](https://aiatlasgraph.com/t/zeroshotcot.md): Instead of examples, a short instruction such as “Let's think step by step” makes the model write out its reasoning.
- [Sieci grafowe (GNN) (Graph neural networks (GNN))](https://aiatlasgraph.com/t/gnn.md): Neural networks operating on graphs (nodes, edges, features), invariant to node ordering; the basis of learning on molecules, social networks and meshes.
  - [Konwolucje grafowe i uwaga na grafach (GCN, GAT) (Graph convolutions and graph attention (GCN, GAT))](https://aiatlasgraph.com/t/messagepassing.md): The most common GNN layers: GCN aggregates a normalised sum of neighbour features, GAT weights neighbours with attention, GraphSAGE samples the neighbourhood.
  - [Sieci geometryczne dla atomów w 3D (równoważność obrotów) (Equivariant graph networks for 3D atomic systems)](https://aiatlasgraph.com/t/equivariantgnn.md): GNNs over atoms with 3D positions, equivariant to rotations, translations and reflections (the E(3) group); they predict energies, forces and molecular properties.
  - [Transformery grafowe (Graph transformers)](https://aiatlasgraph.com/t/graphtransformer.md): Architectures combining a transformer's global attention with graph positional and structural encodings, often alongside local message passing.
- [Sieci konwolucyjne i sekwencyjne (Convolutional and recurrent networks)](https://aiatlasgraph.com/t/cnn.md): Architectures for spatial structure and sequences.
  - [Głębokie architektury konwolucyjne (ResNet, U-Net) (Deep convolutional architectures (ResNet, U-Net))](https://aiatlasgraph.com/t/cnnarchs.md): Families of convolutional architectures: from LeNet, AlexNet and VGG through ResNet with residual connections to U-Net — an encoder-decoder with connections between corresponding levels, used for segmentation and as the backbone of some diffusion models.
  - [LSTM i GRU — sieci z bramkami pamięci (LSTM and GRU (gated recurrent networks))](https://aiatlasgraph.com/t/lstmgru.md): Recurrent networks with gating mechanisms that control how information flows into and remains in hidden state. LSTM uses a separate cell state with input, forget and output gates; GRU combines some of these mechanisms in a simpler design with reset and update gates.
  - [Sieci rekurencyjne (RNN) (Recurrent neural networks (RNN))](https://aiatlasgraph.com/t/rnn.md): Networks that process a sequence step by step while carrying a hidden state from the previous step. Classical RNNs can struggle to learn long dependencies because gradients vanish or explode; LSTMs and GRUs mitigate the problem, and transformers have replaced RNNs in many text tasks.
  - [Splot, filtry i pooling (Convolution, filters and pooling)](https://aiatlasgraph.com/t/convpool.md): The basic operations of a convolutional network: small filters slide over the image and detect local patterns (edges, textures), and pooling reduces feature-map resolution. Filter size, stride and padding determine the output shape.
- [Transformery i attention (Transformers and attention)](https://aiatlasgraph.com/t/transformers.md): Mechanisms for mixing information across a sequence.
  - [Enkoder, dekoder i enkoder–dekoder (BERT, GPT, T5) (Encoder-only, decoder-only and encoder–decoder models)](https://aiatlasgraph.com/t/transformerfamilies.md): The three transformer layouts. Encoders (BERT) read text bidirectionally and serve classification, search and extraction; decoders (GPT, Llama) predict the next token and make up most modern LLMs; encoder–decoders (T5, BART) map one text to another, e.g. translation and summarisation.
  - [FlashAttention](https://aiatlasgraph.com/t/flashattn.md): Attention computed with memory in mind: less data movement, the same result.
  - [Kodowanie pozycji (sinusoidalne, RoPE) (Positional encoding (sinusoidal, RoPE))](https://aiatlasgraph.com/t/posencoding.md): A way to provide token-order information that the attention operation alone does not encode. The original Transformer added sinusoidal vectors; many contemporary LLMs use RoPE, which rotates queries and keys according to position. For context extension RoPE is often additionally scaled or interpolated; some architectures leave part of their layers without explicit positional encoding.
  - [Mechanizm self-attention (zapytania, klucze, wartości) (Self-attention (queries, keys, values))](https://aiatlasgraph.com/t/selfattention.md): The heart of the transformer: each token produces a query (Q), key (K) and value (V), and the output is a weighted average of values, weighted by query–key similarity across all tokens. This lets the model mix information from the whole sequence at once; cost grows quadratically with length. Text-generating models use a causal mask (a token sees only earlier ones).
  - [Oszczędne warianty uwagi (MQA, GQA, MLA) (Memory-efficient attention variants (MQA, GQA, MLA))](https://aiatlasgraph.com/t/attnvariants.md): Multi-head attention variants that reduce KV-cache memory and generation cost: MQA shares one set of keys and values across heads, GQA shares them across groups of heads, and MLA compresses key-value representations into a latent space.

## Sources

- [Stanford CS336 · Language Modeling from Scratch (2025)](<https://cs336.stanford.edu/spring2025/>) `cs336` · reference: Budowa modeli, dane, trening, systemy, inferencja i post-training.
- [Stanford CS231n · Computer Vision (2026)](<https://cs231n.stanford.edu/>) `vision` · reference: Sieci neuronowe i rozumienie obrazów.
- [Stanford CS229 · Machine Learning (2018, archiwum)](<https://cs229.stanford.edu/syllabus-autumn2018.html>) `cs229` · reference: Klasyczne ML, statystyka i podstawy uczenia. Użyte jako stabilny szkielet, nie przegląd nowości.

---

AI Atlas revision 122 (2026-10-30) · HTML: https://aiatlasgraph.com/t/neural · index: https://aiatlasgraph.com/llms.txt · everything: https://aiatlasgraph.com/llms-full.txt · updated 2026-10-30