Neural Network
What is Neural Network?
A neural network is a layered system of mathematical connections that learns patterns from data ( like the brain learning from experience ) and uses those patterns to generate new outputs like images, text, or video.
At a glance
- Also known as
- Artificial neural network (ANN)Deep neural network (DNN)Deep learning model
- Used for
- Learning patterns from large datasets to enable prediction and generationThe foundational architecture for all modern AI generation systemsImage recognition, natural language processing, speech synthesis, generative AI
- Common tools
- TensorFlowPyTorchAll AI generation models (stable diffusion, GPT, CLIP, etc.)
- Related terms
- Deep learningDiffusion modelTransformerGANTrainingParametersModel
- How it works in simple terms
- Data enters the network, passes through layers of interconnected processing units, each of which transforms it in learned ways, and an output emerges from the final layer. During training, the connections are adjusted millions or billions of times to minimise errors. After training, the adjusted connections encode everything the network has learned.
- Where you encounter this
- Every AI generation tool: image generators, video generators, chatbots, voice synthesis systems: is built on one or more neural networks. When you prompt an AI to generate an image or video, you are sending an input through a neural network that transforms it into output.
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
How it compares
Compared with related concepts
Neural networks are often contrasted with classical machine learning and rule-based AI systems. Rule-based systems encode explicit human-written rules for producing outputs; they are transparent and predictable but brittle and unable to handle the complexity and variability of real-world data. Classical machine learning uses hand-engineered features combined with statistical learning. Neural networks, by contrast, learn their own representations directly from raw data without feature engineering, are capable of far greater complexity and nuance, but are also less interpretable and more data-hungry. The success of neural networks over previous approaches is the foundation of the current AI generation revolution.
Think of it like…
A neural network is like a student who learns music not by studying music theory but by listening to thousands of pieces and gradually, through exposure and feedback, developing an internal sense of musical structure, pattern, and style that they can then apply to produce new music that reflects everything they have absorbed.
Pro tip
When a generation model consistently struggles with a specific type of output: unusual cultural aesthetics, rare objects, specific photographic styles: this is often a signal that those elements were underrepresented in the model's training data. Rather than writing longer and more insistent prompts about the same element, try combining the underrepresented element with more common reference points that the model knows well ('traditional Japanese Noh theatre stage, similar spatial arrangement to a Western proscenium theatre') to help the network connect the unfamiliar request to its existing learned representations.
Types and variations
- Convolutional neural networks (CNNs) are specialised for image and spatial data, using localised pattern detection to identify features at multiple scales.
- Recurrent neural networks (RNNs) process sequential data by maintaining a form of memory across time steps.
- Transformers use attention mechanisms to process relationships between all elements of an input simultaneously, making them the dominant architecture for language models and increasingly for image and video generation.
- Generative adversarial networks (GANs) use two competing networks ( a generator and a discriminator ) trained against each other to produce realistic outputs.
- Diffusion models use a specific forward and reverse noise process to generate data.
- Variational autoencoders (VAEs) learn compressed latent representations of data.
- Most modern large-scale AI systems combine multiple architectural types within a single overall system.
Ready to make your first scene in Morphic?
Try MorphicCommon use cases
- Neural networks are the underlying technology for every AI generation tool used in creative production: image synthesis, video generation, natural language generation, speech synthesis, music generation, code generation, image recognition and classification, object detection, translation, summarisation, and virtually every other application of modern machine learning.
- Understanding the neural network as the common foundation of all these systems helps creators understand why similar prompting principles often apply across different modalities and tools: all are variations on the same fundamental learned pattern-recognition architecture.
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
FAQs
A neural network is a computational system composed of layered processing units (neurons) connected by numerical weights, designed to learn patterns from data and produce outputs by applying those learned patterns to new inputs. It is inspired by the structure of biological neural networks and is the foundational architecture behind all modern AI systems, including image generators, language models, and video generation tools.
A neural network learns through a training process in which it is exposed to large quantities of examples with known correct outputs. During training, the weights ( the numerical values of connections between neurons ) are iteratively adjusted through a process called backpropagation to minimise the error between the network's outputs and the correct outputs. After many iterations across many examples, the weights encode learned representations that allow the network to produce accurate outputs on new inputs.
Deep learning refers specifically to neural networks with many hidden layers — 'deep' networks. All deep learning models are neural networks, but not all neural networks are deep learning models (shallow networks with few layers also exist). In practice, the term 'deep learning' is most commonly used for the large, multi-layer networks that power modern AI generation, as their depth is what enables them to learn the complex, hierarchical representations required for high-quality outputs.
Weights are the numerical values of the connections between neurons in a network, which are adjusted during training. Parameters is a broader term that includes weights and biases: all the numerical values in the network that are learned during training. A large language model or image generation model may have billions of parameters, each contributing to the network's overall learned representation. After training, parameters are fixed and define the model's capabilities and characteristics.
A transformer is a specific type of neural network architecture that processes inputs using attention mechanisms, which allow the network to consider relationships between all elements of an input simultaneously rather than sequentially. Transformers are the dominant architecture for language models (GPT, Claude, Gemini) and are increasingly used in image and video generation. They are one of several neural network architectural variants, distinguished by their attention-based processing approach.
Traditional programmes follow explicit rules written by programmers and are limited to what those rules can handle. Neural networks learn their own rules from data, developing internal representations of complex patterns that would be impossible for humans to write explicitly. This is why neural networks can recognise faces in photos, understand the meaning of sentences, and generate photorealistic imagery: tasks that require far more nuance and flexibility than any explicit rule set could provide.
Neural network and artificial intelligence are related but not synonymous. AI is a broad field encompassing many approaches to building systems that exhibit intelligent behaviour. Neural networks are one class of AI architecture: currently the dominant one for most practical AI applications. Earlier AI systems used rule-based approaches, search algorithms, and expert systems rather than neural networks. The current AI generation revolution is specifically a neural network revolution, but AI as a concept is older and broader.
The number of parameters is a rough indicator of a network's representational capacity: how much complexity it can potentially learn. Larger networks generally have greater capacity to represent complex patterns, but also require more data and compute to train effectively. However, size alone does not determine quality: architecture, training data quality, training methodology, and the specific task domain all significantly affect what a model can do. A smaller model trained on excellent domain-specific data can outperform a much larger general model on tasks within that domain.
