Computer Science Machine Learning (from case study)

Bias mitigation

Techniques used to identify and reduce unfair or skewed outputs from a model. Examples: diversifying the training data, auditing outputs for stereotypes, filtering harmful content, reweighting underrepresented groups, and human review.

Character consistency

A generative model's ability to keep the same character's key features (face, clothing, style) across multiple generated images. It is usually achieved with reference images, fixed embeddings, fine-tuning, or fixed random seeds.

Convolutional neural network (CNN)

A neural network that uses convolutional layers, in which small learnable filters (kernels) slide across the input to detect local features such as edges and textures. Pooling layers reduce the data size, and fully connected layers usually produce the final output. It is widely used for image tasks.

Dataset curation

Selecting, cleaning, labelling, balancing and organising training data so it is high quality, relevant and diverse. It includes removing duplicates, errors and harmful content, and it directly affects model accuracy and bias.

Denoising

Removing noise from data to recover the clean underlying signal or image. In diffusion models, the network predicts the noise present at each step and removes it.

Denoising diffusion probabilistic model (DDPM)

A specific diffusion model with two phases. The forward process adds small amounts of Gaussian noise over many steps until the image is pure noise. In the reverse process, a neural network (often a U-Net) learns to predict and remove the noise one step at a time. Training minimises the difference between predicted and actual noise.

Diffusion model

A generative model that learns to reverse a gradual noising process. It generates new images by starting from pure random noise and iteratively denoising it into a coherent image.

Embedding-based approach

Representing inputs (words, images, concepts) as dense numerical vectors in which similar meanings sit close together. In image generation, embeddings (e.g. of a text prompt or a character) guide or condition the model.

Image generation

Creating new images with a model that has learned patterns from training data. It can be conditional or unconditional.

Conditional

Generation guided by extra input, such as a text prompt, class label or reference image, so the output matches that condition.

Unconditional

Generation from random noise alone, with no control over content. The output only resembles the general training data distribution.

Image-to-image translation

Transforming an input image into a different version while preserving its structure. Examples: sketch to photo, day to night, segmentation map to realistic scene.

Noise injection

Deliberately adding random noise to data or model inputs. It is used in the forward process of diffusion models and as input variation in generators, and it can also act as regularisation to improve variety and generalisation.

Segmentation map

An image in which every pixel is labelled with a class (e.g. road, sky, car). It can be the output of segmentation or an input that controls scene layout in image-to-image translation.

Text-to-image generation

Producing an image from a natural-language prompt. The text is converted into an embedding that conditions the generative model.

Training stability

How smoothly and reliably a model's training converges, without oscillating, diverging or failing. GANs are prone to instability, while diffusion models are generally more stable.

Adversarial dynamic

The competition between the generator and the discriminator in a GAN. Each network improves in response to the other, like a minimax game. Ideally training reaches equilibrium, where the discriminator can only guess (about 50% accuracy).

D-dimensional noise vector

A random vector of D values (sampled from a Gaussian or uniform distribution) fed into a GAN generator. Each different vector produces a different output, and D sets the size of the latent space being sampled.

Generative adversarial network (GAN)

A model of two neural networks trained together. A generator creates fake data from noise and a discriminator tries to tell real from fake. Both improve through competition until generated data is hard to distinguish from real data.

Discriminator

The GAN network that classifies input as real (from the training set) or fake (from the generator). Its feedback is used to train the generator.

Generator

The GAN network that takes a noise vector and produces synthetic data (e.g. an image). It is trained to fool the discriminator.

Hybrid model

A model combining two or more generative architectures to use their strengths. Examples: VAE-GAN, and latent diffusion (a VAE compresses images, then diffusion operates in the latent space).

Flow-based model

A generative model that uses a sequence of invertible transformations to map a simple distribution (e.g. Gaussian) to the complex data distribution. Because the mapping is reversible, it can compute exact likelihoods and generate in both directions. Examples: RealNVP, Glow.

Variational autoencoder (VAE)

An autoencoder whose encoder maps input to a probability distribution (mean and variance) in latent space. A sample from that distribution is passed to the decoder, which reconstructs the input. The loss combines reconstruction loss and KL divergence, which gives a smooth, continuous latent space. Outputs tend to be blurrier than GAN outputs.

Latent space

A compressed, lower-dimensional representation of data in which similar items are close together. Sampling or interpolating between points in it generates new data.

Mode collapse

A GAN failure in which the generator produces only a small range of outputs, repeating a few outputs that fool the discriminator, so diversity is lost.