Transformers in Machine Learning

The Transformer architecture, introduced in “Attention Is All You Need” (2017), revolutionized machine learning by replacing recurrence with self-attention. Today, Transformers are the dominant architecture in NLP, computer vision, audio, and increasingly in tabular data.

The Core Problem Transformers Solve

RNNs/LSTMs process sequences step by step — they can’t parallelize training and struggle to connect information from very distant positions. Transformers replace the sequential process with self-attention: every position can directly attend to every other position in a single operation.

Self-Attention

For each token, self-attention computes a weighted sum of all other tokens’ values, where weights are determined by query-key similarity:

Input sequence: "The cat sat on the mat"

For each word:
  Query (Q): "What am I looking for?"
  Key (K):   "What do I contain?"
  Value (V): "What information do I pass on?"

Attention(Q, K, V) = softmax(QKᵀ / √dₖ) × V

The word "sat" attends strongly to "cat" (who sat?) and "mat" (sat where?)

import torch
import torch.nn as nn
import math

class SelfAttention(nn.Module):
    def __init__(self, d_model, num_heads):
        super().__init__()
        self.num_heads = num_heads
        self.d_head = d_model // num_heads

        self.W_q = nn.Linear(d_model, d_model)
        self.W_k = nn.Linear(d_model, d_model)
        self.W_v = nn.Linear(d_model, d_model)
        self.W_o = nn.Linear(d_model, d_model)

    def forward(self, x, mask=None):
        B, T, C = x.shape

        Q = self.W_q(x).view(B, T, self.num_heads, self.d_head).transpose(1, 2)
        K = self.W_k(x).view(B, T, self.num_heads, self.d_head).transpose(1, 2)
        V = self.W_v(x).view(B, T, self.num_heads, self.d_head).transpose(1, 2)

        scores = (Q @ K.transpose(-2, -1)) / math.sqrt(self.d_head)
        if mask is not None:
            scores = scores.masked_fill(mask == 0, -1e9)

        attn = torch.softmax(scores, dim=-1)
        out = attn @ V  # (B, heads, T, d_head)
        out = out.transpose(1, 2).contiguous().view(B, T, C)
        return self.W_o(out)

Transformer Block

A standard Transformer block consists of self-attention + feedforward network with residual connections and layer normalization:

class TransformerBlock(nn.Module):
    def __init__(self, d_model, num_heads, ff_dim, dropout=0.1):
        super().__init__()
        self.attention = SelfAttention(d_model, num_heads)
        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        self.ff = nn.Sequential(
            nn.Linear(d_model, ff_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(ff_dim, d_model),
            nn.Dropout(dropout)
        )

    def forward(self, x, mask=None):
        x = self.norm1(x + self.attention(x, mask))  # Residual + attention
        x = self.norm2(x + self.ff(x))               # Residual + FFN
        return x

Positional Encoding

Unlike RNNs, Transformers process all tokens simultaneously — they have no built-in notion of order. Positional encodings are added to token embeddings to inject position information:

class PositionalEncoding(nn.Module):
    def __init__(self, d_model, max_seq_len=5000):
        super().__init__()
        pe = torch.zeros(max_seq_len, d_model)
        position = torch.arange(max_seq_len).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        self.register_buffer('pe', pe)

    def forward(self, x):
        return x + self.pe[:x.size(1)].unsqueeze(0)

Transformer Families

Encoder-only (BERT-style): Bidirectional attention over the full sequence. Best for understanding tasks: classification, NER, question answering.

Decoder-only (GPT-style): Causal (masked) attention — each token attends only to past tokens. Best for generation: text completion, summarization, code generation.

Encoder-decoder (T5-style): Full encoder + causal decoder. Best for sequence-to-sequence: translation, summarization with fixed-length output.

Using Pretrained Transformers with HuggingFace

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=3)

texts = ["I love this product!", "Terrible experience.", "It's okay."]
inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt", max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = outputs.logits.argmax(dim=-1)

Vision Transformers (ViT)

Transformers now achieve state-of-the-art on image tasks by splitting images into 16×16 patches and treating them as “tokens”:

from torchvision.models import vit_b_16
vit = vit_b_16(pretrained=True)

Transformers are the unified architecture of modern AI — if you understand self-attention, you understand the core mechanism behind GPT, BERT, ViT, Whisper, DALL·E, and most state-of-the-art models.