Varieties of Neural Networks: Architectures, Applications, and Key Concepts

What is the key insight regarding neural network architectures?

Architecture matters just as much as data and training, as each architecture encodes different inductive biases.

What is the primary use case for Fully-connected Neural Networks (FNNs)?

General classification and regression.

What are Convolutional Neural Networks (CNNs) primarily used for?

Image recognition and spatial data.

What do Autoencoders (AEs) learn to do?

Compress data into a small representation and then reconstruct it.

What is the core mechanism of Recurrent Neural Networks (RNNs)?

Recurrence, where outputs from one time step feed back as inputs to the next, creating a form of memory.

What is the primary function of Generative Adversarial Networks (GANs)?

To generate realistic images through adversarial training of two networks: a generator and a discriminator.

What is the role of the generator in a GAN?

To take random noise as input and produce realistic fake images.

What is the role of the discriminator in a GAN?

To classify real images and fake images, determining which is which.

What is the core idea behind Latent Diffusion Models (LDMs)?

To combine autoencoders with a denoising process to generate high-quality images from text prompts.

What is the first step in the LDM process?

Train an autoencoder on a large image dataset to create a compact latent space.

What is the purpose of adding noise in the LDM process?

To corrupt the compressed latent representation and train a denoising network to predict and remove the noise.

What does the denoising network do in the LDM process?

It takes pure random noise and iteratively 'denoises' it into a valid latent representation.

What is the significance of the latent space in Autoencoders?

It is the most information-dense part, where data is compressed.

What is a key property of Autoencoders?

They are a form of unsupervised learning, requiring no labeled data.

What is a major weakness of Recurrent Neural Networks (RNNs)?

They suffer from vanishing gradients, making it difficult to retain information from many steps back.

What is the main advantage of Variational Autoencoders (VAEs) over standard Autoencoders?

VAEs have a probabilistic latent space, allowing for smoother generation of new data points.

What is the main application of GANs in the real world?

They can generate photorealistic human faces and have raised concerns about deepfakes.

What is the training process of GANs characterized by?

A generator that improves by trying to fool a discriminator, which simultaneously learns to detect fakes.

What type of data do Autoencoders struggle to generalize well outside of?

They do not generalize well outside of the training data, which is useful for anomaly detection.

What is a common use case for RNNs?

Language understanding, where the meaning of words depends on prior words.

What analogy is used to describe the function of Autoencoders?

Like zipping a file (encoder) and then unzipping it (decoder).

What is the defining feature of RNNs?

Their ability to track context across sequences through recurrence.

What is the impact of GANs on image generation technology?

They have progressed from generating blurry low-res images to creating photorealistic outputs.

What is the primary use case for Deep Reinforcement Learning (DRL)?

Game playing and sequential decision-making.

What is the main purpose of Transformers in neural networks?

To serve as language models and support modern AI applications.

What is the purpose of the Encoder in a Variational Autoencoder (VAE)?

To compress images into latent space.

What does the Denoising Network (U-Net) do in the context of image processing?

Learns to reverse noise, guided by text.

What is the core idea of Deep Reinforcement Learning (DRL)?

An agent learns by interacting with an environment, receiving rewards for good actions and penalties for bad ones.

What landmark paper introduced the Deep Q-Network (DQN)?

'Human-level control through deep reinforcement learning' by Mnih et al., 2015.

What type of data does the DQN process?

Raw game screen pixels (stacked frames for temporal information).

What was a significant performance achievement of DQN?

Achieved human-level or above performance on many Atari games.

What is a key limitation of DQN in complex games?

It has no memory, making it struggle with games requiring long-term planning.

What is the purpose of AlphaGo's Rollout policy network?

Makes quick move predictions using a fast, shallow network.

What does the Value network in AlphaGo estimate?

The probability of winning from any board position.

What is the theoretical claim made in 'Reward is Enough' by Silver et al.?

Intelligence can emerge from a single goal — maximizing reward.

What distinguishes Supervised Learning from Unsupervised Learning?

Supervised learning requires labeled data, while unsupervised learning finds structure in raw data.

What is the difference between Generative and Discriminative Models?

Generative models create new data samples, while discriminative models classify or distinguish data.

Define Latent Space in the context of neural networks.

A compressed, lower-dimensional representation of data used for internal encoding.

What are the components of Reinforcement Learning?

Agent, Environment, State, Action, Reward, Policy.

What is the key difference between a standard autoencoder and a variational autoencoder?

The VAE incorporates a probabilistic latent space for generating new data.

Why can an autoencoder be effective for anomaly detection?

It learns to reconstruct normal data, making anomalies easier to identify when reconstruction fails.

What is the main architectural difference between GANs and VAEs?

GANs use a generator-discriminator framework, while VAEs use an encoder-bottleneck-decoder structure.

What role does self-play have in AlphaGo's training?

It allows the RL policy network to improve by playing against itself rather than just imitating human moves.

What is the primary use of Convolutional Neural Networks (CNNs)?

Image recognition.

What is the significance of inductive bias in machine learning models?

It allows models to generalize beyond the training data, influencing learning efficiency.

What is the primary use of Recurrent Neural Networks (RNNs)?

Processing sequential or language data.

What is the main advantage of using Latent Diffusion Models (LDMs)?

They perform denoising in latent space, making them computationally efficient for tasks like text-to-image generation.

What happens to the generator and discriminator in a GAN during training?

They compete against each other, improving until they reach a state of equilibrium.

What is a common application for Autoencoders?

Compression, representation learning, and anomaly detection.

What is the difference in training signals between Supervised Learning and Reinforcement Learning?

Supervised learning uses labeled input-output pairs, while reinforcement learning uses rewards and penalties based on actions taken.