Welcome to the second set of flashcards designed to help you build or revise your machine learning (ML) knowledge whenever needed. If you missed the first set covering Reinforcement Learning fundamentals, it's a great place to start. This set first appeared in October 2024, just after the Nobel Prizes in Physics and Chemistry recognized work closely connected to modern machine learning. John Hopfield and Geoffrey Hinton were honored for foundations of neural-network learning; Demis Hassabis and John Jumper for protein-structure prediction, alongside David Baker for computational protein design. That’s why these flashcards dive into this crucial subset of machine learning.
As a mother of five, I always think about how I would explain these concepts to my kids. So you’ll find two explanations: one for adults and one for “explain me as if I was five”.
In these cards, we walk through the main types of deep learning you are likely to meet in modern ML: feedforward neural networks (FNNs), multilayer perceptrons (MLPs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), generative adversarial networks (GANs), autoencoders and variational autoencoders (VAEs), diffusion models, and transformer models with self-attention. The point is not to memorize every acronym. It is to understand what kind of data each architecture was built to handle, what problem it made easier, and why many newer AI systems combine several of these ideas instead of treating them as isolated families.
What Is Deep Learning?
Deep learning is a branch of machine learning that uses neural networks with multiple layers to learn useful representations directly from data. Different architectures place different assumptions on that learning: convolutions favor local patterns, recurrence carries state through sequences, attention connects distant elements, and generative models learn how examples are distributed.
Nine Types of Deep Learning at a Glance
Type | Best suited to | Typical job |
|---|---|---|
Feedforward neural network (FNN) | Fixed-size feature vectors | General classification and regression |
Multilayer perceptron (MLP) | Tabular or vectorized data | Learning nonlinear mappings |
Convolutional neural network (CNN) | Images, video frames, spatial grids | Detecting local visual patterns |
Recurrent neural network (RNN) | Ordered sequences and streams | Modeling step-by-step dependencies |
Long short-term memory (LSTM) | Longer sequences with delayed signals | Retaining information across many steps |
Generative adversarial network (GAN) | Images and other sample-based data | Generating realistic examples |
Autoencoder / VAE | High-dimensional data with latent structure | Compression, representation learning, generation |
Diffusion model | Images, audio, video and other continuous signals | High-quality generation through iterative denoising |
Transformer | Text, code, image patches and multimodal tokens | Connecting distant elements with attention |
FAQ
What are the main types of deep learning?
Common families include feedforward networks and MLPs, CNNs, RNNs and LSTMs, transformers, autoencoders and VAEs, GANs, and diffusion models. This is a practical map rather than a strict taxonomy: modern systems often combine several of these architectures.
Is a multilayer perceptron the same as a feedforward neural network?
An MLP is a common kind of feedforward neural network. “Feedforward” describes the direction of computation: information moves from input to output without a recurrent loop. An MLP usually means a feedforward network with one or more hidden layers and nonlinear activation functions.
When should you use a CNN instead of a transformer?
CNNs remain a strong choice when local spatial patterns, efficiency, and limited training data matter. Transformers are useful when a model must connect distant parts of an input or combine multiple modalities. Many modern vision systems mix convolution and attention rather than treating them as exclusive choices.
What is the difference between an RNN and an LSTM?
An RNN carries a hidden state from one step to the next, but standard RNNs can struggle to preserve signals across long sequences. An LSTM adds gates that control what to remember, update, and forget, making longer dependencies easier to learn.
Are GANs, VAEs, and diffusion models all generative models?
Yes, but they learn and generate differently. GANs train a generator against a discriminator. VAEs learn a structured latent space and decode samples from it. Diffusion models learn to reverse a gradual noising process, usually trading slower generation for stable training and high output quality.
Nobel Prize part
You are welcome to share these cards! For our Premium subscribers, we prepared a downloadable PDF-version and a collection of resources to dive deeper (I’m very proud of it)→
📚 AI 101 Concepts series: #1 Reinforcement Learning & Deep Learning · #2 Types of Deep Learning · #3 RLHF, RLAIF, RLEF, RLCF · #4 Supervised & Unsupervised Learning · #5 AI Learning Techniques
You can get it too →
The flashcards series is a pure experiment, and it will be evolving with time and your feedback. If you want to help create them or can recommend cool tools that could assist with this – let me know at [email protected]
How did you like it?
We might contact you to ask for more feedback. Thank you!
