Tag
Dropout is a regularization technique that switches neurons off at random during training, with retention probability p, and divides the surviving activations by p to preserve their scale. At inference the full network runs without dropping anything. Mathematically it amounts to averaging a huge ensemble of subnetworks that share weights.
Layer normalization stabilizes training by normalizing each example on its own, using the mean and standard deviation of its own activations. Unlike batch normalization, it does not depend on the batch size, which is why transformers adopted it. RMSNorm is its lighter, most widely used variant today in large language models.
Batch normalization is a technique that normalises each layer's activations using the mean and variance of the mini-batch, then rescales them with two learnable parameters, gamma and beta. Introduced in 2015, it enables higher learning rates, speeds up training and stabilises deep neural networks during optimisation.
Weight initialization sets the starting scale of the matrix W before training begins. Xavier/Glorot, from 2010, splits the variance between inputs and outputs and suits sigmoid and tanh; He, from 2015, doubles it for ReLU, which zeroes half the activations. A poor choice stalls or breaks learning.
Second-order methods use the Hessian, the matrix of second derivatives of the loss, to orient each step better than the gradient alone. Newton's method updates the weights by subtracting the inverse Hessian times the gradient, but inverting it is far too costly in large networks, so practitioners fall back on quasi-Newton approximations.
Adam combines a first-order moment (a moving average of the gradient) and a second-order moment (an average of its squares) to give each parameter its own learning rate. With bias correction and the defaults β1=0.9, β2=0.999 and ε=1e-8, it converges fast with almost no tuning.
Momentum is an improvement to gradient descent that accumulates a velocity from past gradients instead of looking only at the current one. With a coefficient β around 0.9, that inertia smooths the zigzag of plain SGD, pushes through long narrow valleys and makes training converge noticeably faster than gradient descent alone.
In large neural networks the loss surface almost never traps training in a bad local minimum. The real obstacle is saddle points, far more common in high dimensions. This article explains local versus global minima, why stochastic gradient descent slips past them, and what role convexity plays in the whole picture of optimisation.
Backpropagation shares out the blame for the error among all the weights of a neural network. It propagates an error signal backwards layer by layer, multiplying by the local derivatives, and so obtains the gradient of every weight in a single pass. That idea, published in 1986, is what makes training deep networks possible.
Gradient descent is the algorithm that trains almost every neural network. It computes the slope of the loss function with respect to each weight and takes a small step in the opposite direction, controlled by the learning rate, until it reaches a minimum where the error stops falling and the model has converged.
Mean absolute error (MAE) averages the absolute difference between prediction and reality, so a single outlier weighs only its fair share and never blows up the loss. Huber loss combines that robustness with the smoothness of mean squared error through a delta parameter that decides where the behaviour switches over.
A loss function measures how wrong a neural network is on a single example, comparing its prediction with the correct value. The cost function averages that loss across the whole dataset. That single number is exactly what training tries to reduce, step by step, using gradient descent to adjust every weight.
Forward propagation is the process by which a neural network turns its input into a prediction, one layer at a time. Each layer multiplies the input vector by a weight matrix, adds a bias and applies an activation function, so the output of one layer feeds the next until the final result appears at the end.
The exploding gradient problem happens when the gradient norm grows without control during backpropagation, especially in deep and recurrent networks with large weights. Training destabilises and the loss turns into NaN. Gradient clipping, together with good initialisation and normalisation, is the standard fix used today.
The vanishing gradient problem appears when the error signal shrinks as it backpropagates through many layers. The sigmoid and hyperbolic tangent functions have small derivatives, and their product tends toward zero, so the early layers barely learn. ReLU, careful initialisation and normalisation solve the problem.
Choosing an activation function is simple with one base rule: use ReLU in the hidden layers, GELU or SiLU in transformers, and reserve the output for softmax in multiclass classification, sigmoid in binary problems and a linear activation in regression. This comparison gathers formulas, ranges and use cases.
The Mish function is defined as mish(x) = x·tanh(softplus(x)). It is smooth, non-monotonic and self-regularized, properties that let it outperform Swish and ReLU across many tests. The YOLOv4 object detector adopted it in its backbone as the default activation.
The Softplus function is defined as softplus(x) = ln(1 + eˣ): a smooth, always-positive approximation of ReLU whose derivative is exactly the sigmoid. It is differentiable across its whole domain, avoids ReLU's angular kink and its output never quite reaches zero.
The SELU (Scaled Exponential Linear Unit) function is defined as SELU(x) equal to lambda times ELU(x), with lambda near 1.0507 and alpha near 1.6733. Those two constants make activations converge on their own towards zero mean and unit variance layer after layer, building deep networks that normalize themselves without batch normalization.
The ELU (Exponential Linear Unit) activation function returns x for positive inputs and α(eˣ−1) for negative ones. By allowing smooth negative values, it pushes the mean of the activations toward zero, speeds up convergence compared with ReLU, and avoids dead neurons thanks to a gradient that never vanishes completely on the negative side.
The Swish function, also called SiLU, multiplies the input by its sigmoid: swish(x) = x·σ(x). It is smooth, non-monotonic and self-gating, so it often beats ReLU in deep networks. Models like LLaMA and EfficientNet use it as their default activation.
The GELU (Gaussian Error Linear Unit) function multiplies each input by the probability that a standard normal falls below that input. The result is a smooth curve with a continuous derivative that weights inputs by their magnitude, and it has become the default activation inside BERT and GPT.
An activation function is the nonlinear operation each neuron applies to its weighted sum z to produce its output a equals f of z. Without it, stacking layers only chains linear transformations and the whole network collapses into a single one. That nonlinearity is what lets a network learn complex patterns.
In an artificial neuron, weights measure how important each input is and the bias shifts the result. The neuron multiplies each input by its weight, adds everything up and includes the bias to produce the weighted sum z = Wx + b, the number that then passes through the activation function.
The exponential function eˣ and its inverse, the natural logarithm ln(x), appear again and again in deep learning. The exponential builds the sigmoid and softmax that turn numbers into probabilities, while the natural logarithm defines cross-entropy, the loss used to train almost every classifier in practice today.
Matrix multiplication is the core operation of a neural network: each layer gathers its weights into a matrix W and computes its output as the product W times X. That single operation, repeated layer after layer, turns the inputs into predictions and explains why graphics cards dominate modern deep learning.
The mathematics of neural networks rests on three blocks: linear algebra represents data and weights as vectors and matrices, calculus with derivatives and the chain rule lets the network learn through gradient descent, and probability shapes the loss functions. This roadmap walks that path from beginning to end so you know what to study and in what order.
DINOv2 is Meta AI's computer vision model, trained via self-supervision on 142 million images with no human labels. With a simple linear layer on the frozen encoder, it matches or beats supervised models on ImageNet classification, semantic segmentation and monocular depth estimation.
The hyperbolic tangent (tanh) is an activation function that squashes any real value into the interval (-1, 1) with zero-centred output, which removes the systematic gradient bias seen with sigmoid. It is the standard activation inside the LSTM and GRU memory cells used in recurrent networks.
The sigmoid function maps any real number to a value between 0 and 1, which makes it the natural activation function for expressing probabilities in a neural network. It is differentiable across its whole domain, though it suffers from saturation and vanishing gradients in deep networks, so today it is mostly reserved for the output layer.
The Softmax function transforms a neural network's output logits into a probability distribution that sums to 1. It is the standard activation for multi-class classification, from image classifiers to the vocabulary layer of language models like GPT.
Leaky ReLU is an activation function derived from ReLU that replaces the zero output for negative inputs with a small slope, typically 0.01. This keeps the gradient from ever reaching zero, which prevents the dying neuron problem and stabilizes training in very deep convolutional, recurrent, and GAN networks.
ReLU (Rectified Linear Unit, f(x) = max(0, x)) has been the dominant activation function in deep learning since AlexNet popularised it in 2012: cheap to compute, resistant to vanishing gradients, and with one well-known weakness, dying ReLU.
The step function, or Heaviside function, is the simplest activation function in a neural network: it maps any input to a binary output, 0 or 1, depending on whether it crosses a fixed threshold. It was the core decision mechanism of Rosenblatt's perceptron in 1958, but its derivative is zero almost everywhere, so modern networks use sigmoid or ReLU instead.
The linear function f(x) = ax + b is the simplest activation a neural network can use: in its identity form, f(x) = x, it is the standard choice for the output layer in regression, because it does not bound the range of possible values. In hidden layers it fails, because composing several linear functions collapses the whole network into one equivalent linear layer.
A fully connected neural network, also called a dense network, is the fundamental architecture of deep learning: every neuron in a layer connects to all neurons in the previous and next layer. This total connectivity lets it approximate any continuous function, though its computational cost grows quadratically with the number of neurons.
In a neural network, the input is represented as a column vector x in R^n that the hidden layer transforms through a weight matrix W, a bias vector b, and a non-linear activation function such as ReLU, sigmoid, or tanh. Training adjusts W and b by minimising the loss function via gradient descent and backpropagation.
A multilayer neural network consists of an input layer, one or more hidden layers, and an output layer, where each neuron weights its inputs and applies a non-linear activation function before passing the result to the next layer. Through forward propagation and backpropagation, the network adjusts millions of weights to learn hierarchical representations capable of classifying images, translating text, or generating language.
Transfer learning lets you reuse a model already trained on a massive dataset, such as ImageNet or a large text corpus, to solve a new task with far less proprietary data and compute time. It works through fine-tuning, feature extraction, or prompting, and it performs best when the source and target domains are similar to each other.
Adversarial machine learning studies deliberate attacks on AI systems (evasion, poisoning and model extraction) and the defenses used to resist them, chiefly adversarial training, robustness certification and monitoring the distribution of production input data.
Federated learning trains AI models collaboratively across many devices or organisations without moving the original data: each participant trains locally and sends only gradients to the central server. Formalised by Google in 2016, it does not guarantee privacy on its own: it needs differential privacy or secure aggregation to prevent leaks from those gradients.
Reinforcement learning is the AI technique in which an agent learns to make optimal decisions through trial and error, without labelled data: it acts in an environment, receives a reward or penalty based on the outcome, and adjusts its strategy to maximise long-term cumulative reward.
Computer vision is the branch of artificial intelligence that lets machines interpret digital images: detecting objects, segmenting regions and recognising patterns through convolutional neural networks. Since 2012, when AlexNet cut ImageNet classification error to 15.3%, it has spread into manufacturing, medicine, transport and precision agriculture.
Deep neural networks are today the foundation of almost every artificial intelligence application: from facial recognition to machine translation. Built on architectures like CNNs, RNNs, and Transformers, deep learning has transformed computer vision, speech recognition, and natural language processing over the last decade.
Modern artificial intelligence rests on three pillars: machine learning, deep neural networks, and natural language processing. These techniques have pushed image recognition and machine translation past human-level precision on specific tasks, though the overall system still depends on quality data and constant human oversight.
