Having explored the sinusoidal positional encoding scheme, which injects sequence order information using fixed trigonometric functions, we now turn our attention to an alternative approach: learned positional embeddings. Both methods aim to solve the same problem – making the permutation-invariant self-attention mechanism aware of sequence order – but they do so with different characteristics and trade-offs. Understanding these differences is important for choosing the right approach for a given task or model architecture.

Let's systematically compare these two techniques across several dimensions.

Parameter Efficiency and Model Size

Sinusoidal Positional Encodings: This method is parameter-free. The encoding vectors are generated using a predefined formula: $PE_{(pos, 2i)} = \sin(pos / 10000^{2i/d_{model}})$ $PE_{(pos, 2i+1)} = \cos(pos / 10000^{2i/d_{model}})$ These values can be pre-computed up to a maximum expected sequence length and stored, or calculated on-the-fly during the forward pass. In either case, they do not add any trainable parameters to the model. This makes them attractive for models where parameter count is a constraint.

Learned Positional Embeddings: This approach treats positions as discrete tokens. An embedding matrix is created, typically of size $L_{max} \times d_{model}$ , where $L_{max}$ is the maximum sequence length the model is designed to handle and $d_{model}$ is the embedding dimension (matching the token embeddings). Each row $j$ in this matrix is a learnable vector representing position $j$ . During the forward pass, the appropriate positional embedding vector is looked up based on the position index and added to the corresponding token embedding. This introduces $L_{max} \times d_{model}$ additional trainable parameters to the model. For models with very large context windows (large $L_{max}$ ) or high dimensionality ( $d_{model}$ ), this can represent a non-trivial increase in model size and memory requirements.

Flexibility and Inductive Bias

Sinusoidal Positional Encodings: The fixed nature of sinusoidal encodings introduces a strong inductive bias into the model. The trigonometric functions inherently encode positions in a way that facilitates modeling relative positions. Because of trigonometric identities like $\sin(\alpha+\beta)$ and $\cos(\alpha+\beta)$ , the encoding $PE_{pos+k}$ for a position offset by $k$ can be expressed as a linear transformation of $PE_{pos}$ . This structure potentially makes it easier for the self-attention mechanism to learn relationships based on relative offsets, independent of absolute position. However, this fixed structure might not be optimal for all tasks or data distributions.

Learned Positional Embeddings: Being learnable, these embeddings offer maximum flexibility. The model can, in principle, learn any positional representation that minimizes the training loss for the target task. This allows the model to adapt the positional information specifically to the nuances of the data. However, this flexibility comes at the cost of a weaker inductive bias. Without the inherent structure of sinusoidal functions, the model must learn positional relationships entirely from the data, which might require more data or training time, especially for capturing complex relative positioning patterns. There's also a risk that the learned embeddings might not discover as smooth or generalizable a representation as the sinusoidal functions provide.

Extrapolation to Unseen Sequence Lengths

Sinusoidal Positional Encodings: The mathematical formulation allows sinusoidal encodings to be generated for any position index, even those beyond the maximum length encountered during training. This suggests a potential advantage in extrapolating to longer sequences. While the model's attention mechanism itself wasn't trained on these longer interactions, the positional signals remain well-defined. The practical effectiveness of this extrapolation can vary, as model performance might still degrade on significantly longer sequences due to distribution shift, but the encoding mechanism itself does not break down.

Learned Positional Embeddings: This method typically struggles with extrapolation. Since embeddings are only learned for positions up to $L_{max}$ , the model has no defined representation for positions $L_{max}, L_{max}+1, \dots$ . Simply extending the sequence during inference would require either assigning zero vectors, reusing existing vectors, or employing more sophisticated techniques like embedding interpolation or retraining/fine-tuning the model with longer sequences. Standard implementations often fail or produce unpredictable results when presented with sequences longer than $L_{max}$ .

Empirical Performance

Sinusoidal Positional Encodings: Used in the original "Attention Is All You Need" Transformer, they demonstrated strong performance on machine translation tasks. They often provide a solid baseline and perform well across various sequence modeling problems.

Learned Positional Embeddings: Adopted by many influential models like BERT and the GPT series. Empirical results show that learned embeddings can achieve state-of-the-art performance, suggesting that the flexibility they offer allows models to capture positional information effectively, given sufficient data and model capacity.

In practice, for sequence lengths commonly encountered during training, the performance difference between the two methods is often small. The choice might hinge more on other factors like parameter budget, desired extrapolation capabilities, or specific architectural choices (e.g., relative positional encodings, discussed later, modify this comparison).

Implementation Complexity

Both methods are relatively straightforward to implement in modern deep learning frameworks.

Sinusoidal: Requires implementing the mathematical formula, often involving matrix operations for efficiency. Care must be taken with numerical stability and correctly broadcasting the encodings to match the input batch dimensions.

Learned: Often simpler conceptually, typically requiring only the instantiation of an embedding layer (e.g., torch.nn.Embedding or tf.keras.layers.Embedding) and adding its output to the token embeddings.

Summary Table

Feature	Sinusoidal Positional Encoding	Learned Positional Embedding
Parameters	None (Fixed function)	Adds $L_{max} \times d_{model}$ parameters
Flexibility	Lower (Fixed structure)	Higher (Learned from data)
Inductive Bias	Strong (Encourages relative positioning)	Weaker (Must learn positional relations from scratch)
Extrapolation	Theoretically possible, generates valid encodings	Poor, requires specific handling for unseen positions
Max Length Constraint	Defined by computation, not parameters	Hard constraint based on $L_{max}$ during training
Data Requirement	Potentially less data-hungry due to built-in structure	May require more data to learn effective representations
Notable Usage	Original Transformer	BERT, GPT series
Implementation	Requires implementing the formula	Often involves a standard Embedding layer

Conclusion

The choice between sinusoidal and learned positional embeddings involves a trade-off between inductive bias, parameter efficiency, flexibility, and extrapolation capabilities. Sinusoidal encodings offer a parameter-free, structured way to represent position with good theoretical properties for relative positioning and extrapolation. Learned embeddings provide greater flexibility, allowing the model to tailor positional representations to the specific task, but they add parameters and struggle with sequences longer than those seen during training. The optimal choice often depends on the specific requirements of the model, the scale of the data, and the desired sequence length capabilities. As we will see in later chapters, developments like relative positional encodings aim to combine the benefits of structured positional awareness with context-dependent representations.