Positional Encoding
Positional encoding adds information about token position to input embeddings in Transformer models, which otherwise have no inherent sense of sequence order. This enables the model to understand word order and sentence structure.
Understanding Positional Encoding
Positional encoding is a technique used in transformer architectures to inject information about the order and position of tokens in a sequence, since the self-attention mechanism itself is permutation-invariant and has no inherent notion of sequence order. The original transformer paper introduced sinusoidal positional encodings that use sine and cosine functions of different frequencies to represent each position, allowing the model to learn relative positioning patterns. Without positional encoding, a transformer would treat "the cat sat on the mat" identically to any rearrangement of those same words. Modern variants include learned positional embeddings, where position representations are trained alongside other parameters, and rotary position embeddings (RoPE) used in models like LLaMA. Positional encoding is essential for tasks where word order carries meaning, which encompasses virtually all natural language processing and sequence modeling applications.
Category
Deep Learning
Is AI recommending your brand?
Find out if ChatGPT, Perplexity, and Gemini mention you when people search your industry.
Check your brand — $9Related Deep Learning Terms
Activation Function
An activation function is a mathematical function applied to a neuron's output to introduce non-linearity into a neural network. Common activation functions include ReLU, sigmoid, and tanh, each with different properties for gradient flow.
Adam Optimizer
Adam (Adaptive Moment Estimation) is an optimization algorithm that combines the benefits of AdaGrad and RMSProp. It adapts learning rates for each parameter using estimates of first and second moments of gradients.
Adapter Layers
Adapter layers are small trainable modules inserted into a pre-trained model to enable parameter-efficient fine-tuning. They allow task adaptation while keeping the original model weights frozen.
Attention Mechanism
An attention mechanism allows neural networks to focus on the most relevant parts of the input when producing each element of the output. Attention is the foundational innovation behind the Transformer architecture and modern large language models.
Autoencoder
An autoencoder is a neural network trained to compress input data into a compact representation and then reconstruct it. Autoencoders are used for dimensionality reduction, denoising, and learning latent representations.
Backpropagation
Backpropagation is the algorithm used to train neural networks by computing gradients of the loss function with respect to each weight. It propagates error signals backward through the network to update weights and minimize prediction errors.
Batch Normalization
Batch normalization is a technique that normalizes layer inputs across mini-batches during training to stabilize and accelerate neural network training. It reduces internal covariate shift and allows higher learning rates.
Batch Size
Batch size is the number of training examples used in one iteration of gradient descent. Larger batches provide more stable gradient estimates but require more memory, while smaller batches add beneficial noise.