🤖 AI / ML

Deep Learning

deep learning · dl · neural networks · deep neural networks · neural nets · deep-learning

Definition

The depth of a neural network is its number of hidden layers. Each layer extracts increasingly abstract features: the first captures edges and textures in an image, the last captures high-level concepts. Deep networks demand large volumes of data and GPU compute, but outperform classical ML methods on complex tasks.

The architecture of a neural network

Input layer       Hidden layers (deep)        Output layer
──────────        ────────────────────        ─────────────
User features  →  Layer 1: basic              Product
Item features  →  patterns              →     recommendation
Context        →  Layer 2: combinations  →    (probability of
               →  Layer N: high-level    →    click or purchase)
                  concepts

Deep learning in e-commerce: applications

Task	Architecture	Example
Recommendations	Two-tower, NCF, DLRM	Frequently bought together
Semantic search	BERT, bi-encoder	“green dress for an evening out”
AI shopping assistant	LLM plus RAG	Answering questions about products
Visual search	CNN, Vision Transformer	Finding a product from a photo
Churn prediction	LSTM, Transformer	Predicting churn from an event sequence

Deep learning versus classical ML

Dimension	Classical ML	Deep learning
Feature engineering	Manual	Automatic
Data volume	Works on small samples	Needs large volumes
Interpretability	High	Low (a black box)
Training speed	Fast	Slow, GPU required
Inference	Fast	Depends on model size
Preferred when	Tabular data, modest volume	Text, images, large volume

The two-tower architecture for recommendations

The most widely deployed deep learning architecture for production recommender systems:

User tower: user ID, purchase history, affinity  → user embedding (128-dim)
Item tower: item ID, category, attributes, price → item embedding (128-dim)

Score = dot_product(user_emb, item_emb)
The top-K items by score become the recommendations

The advantage is operational: the item tower is precomputed offline, and the top-K lookup runs in
milliseconds through an ANN index such as FAISS or ScaNN.

Frequently asked questions

How is deep learning different from classical machine learning?

Classical algorithms such as Random Forest and XGBoost rely on manual feature engineering: an expert constructs features from the data. Deep learning extracts features automatically through the layers of a network. It outperforms classical methods on unstructured data — text, images, audio — but needs more data and more compute.

How is deep learning used in recommender systems?

Deep learning makes three things possible: building user and item embeddings in one vector space so similar objects can be found; modelling complex non-linear relationships between behaviour and preference; and processing sequences of events with transformer-based models to predict the next purchase.

How do neural recommendations compare with classical matrix factorization?

Matrix factorization (ALS, SVD) is a linear model: the interaction matrix is decomposed into latent factors. Neural recommenders (NCF, DLRM, two-tower) are non-linear and capture complex feature interactions. In practice neural models are more accurate on large data sets, but require GPU infrastructure and more demanding operations.