Recurrent Neural Networks and Sequence Models

Recurrent Neural Networks and Sequence Models

Text doesn't come all at once. When you read this sentence, you process it word by word, remembering context. "The bank executive said the merger was...[wait, was the bank or the river involved?]" Order matters. Recurrent Neural Networks (RNNs) are designed for this: processing sequences where previous inputs matter.

Part 1: Why CNNs Aren't Enough—Sequences Need Memory

CNN: excellent for spatial patterns (images). But text, speech, time series are sequential. Word order is critical.

Example: Sentiment analysis

# CNN approach (treating words as pixels): FAILS
# "I loved the movie" → sentiment = positive ✓
# "The movie was bad" → sentiment = negative ✓
# "I would not say the movie was bad" → sentiment = ???
# CNN might miss that "not" negates "bad" because it looks at local patterns

# RNN approach: SUCCEEDS
# Process word by word, maintaining context
# "not" → "say" → "bad" → "was" → "movie"
# RNN learns that "not ... bad" usually means positive sentiment

The RNN Idea: At each time step, pass the previous hidden state forward. This "hidden state" is the network's memory.

Part 2: RNN Architecture and Forward Pass

import numpy as np

# Simplified RNN for text: character-level prediction
# Goal: given "hel", predict next character is "l" (spell "hello")

class SimpleRNN:
    def __init__(self, vocab_size, hidden_size):
        # vocab_size = number of unique characters (26 letters + space, etc.)
        # hidden_size = memory size (typically 128, 256, 512)

        # Input → Hidden
        self.Wxh = np.random.randn(vocab_size, hidden_size) * 0.01
        # Hidden → Hidden
        self.Whh = np.random.randn(hidden_size, hidden_size) * 0.01
        # Hidden → Output
        self.Why = np.random.randn(hidden_size, vocab_size) * 0.01

        self.bh = np.zeros((1, hidden_size))
        self.by = np.zeros((1, vocab_size))

    def forward(self, inputs, h_prev=None):
        """
        inputs: sequence of one-hot encoded characters
        h_prev: hidden state from previous sequences (initially None)
        """
        if h_prev is None:
            h_prev = np.zeros((1, self.Whh.shape[0]))

        xs, hs, ys = {}, {}, {}
        hs[-1] = np.copy(h_prev)

        # Forward pass through time
        for t in range(len(inputs)):
            xs[t] = inputs[t]
            # Hidden state at time t
            hs[t] = np.tanh(np.dot(xs[t], self.Wxh) + np.dot(hs[t-1], self.Whh) + self.bh)
            # Output at time t
            ys[t] = np.dot(hs[t], self.Why) + self.by

        return xs, hs, ys

# Example sequence: "h", "e", "l"
# After processing:
# - Hidden state h_0 captures: "after seeing 'h'"
# - Hidden state h_1 captures: "after seeing 'h' then 'e'"
# - Hidden state h_2 captures: "after seeing 'h', 'e', then 'l'"

# This hidden state is the MEMORY!

Key Equation (Hidden State Update):

hₜ = tanh(Wxh · xₜ + Whh · hₜ₋₁ + bh)

• Wxh · xₜ: new information from current input
• Whh · hₜ₋₁: memory from previous step
• tanh: activation function (squashes to [-1, 1])

Part 3: The Vanishing Gradient Problem

Problem: Long sequences forget distant context.

# Example: "The student studied hard for the JEE exam and scored 99 percentile.
# Despite the success, the student's family was proud."

# To predict "proud", the RNN needs to remember "success" from many words back.
# But through backpropagation, gradients exponentially decay.

# Chain rule through time:
# ∂loss/∂W = ∂loss/∂h_t · ∂h_t/∂h_{t-1} · ∂h_{t-1}/∂h_{t-2} · ... · ∂h_1/∂W

# If each ∂h_t/∂h_{t-1} is 0.8, after 100 steps: 0.8^100 ≈ 10^-10 (vanishes!)
# If each is 1.2, after 100 steps: 1.2^100 ≈ 10^10 (explodes!)

# This is why simple RNNs can't learn long-range dependencies

Part 4: LSTM—The Solution

LSTM (Long Short-Term Memory) uses gates to control information flow. Instead of one hidden state, it maintains a cell state (long-term memory) separate from hidden state (short-term output).

# LSTM has 3 gates: forget, input, output

class LSTMCell:
    def __init__(self, input_size, hidden_size):
        self.hidden_size = hidden_size

        # Weights for all 4 inputs (3 gates + candidate)
        # Each gate needs weights from input and previous hidden state
        self.W = np.random.randn(input_size + hidden_size, 4 * hidden_size) * 0.01
        self.b = np.zeros((1, 4 * hidden_size))

    def forward(self, x, h_prev, c_prev):
        """
        x: current input
        h_prev: previous hidden state
        c_prev: previous cell state (long-term memory)
        """
        # Concatenate input and previous hidden state
        combined = np.concatenate((x, h_prev), axis=1)

        # Compute all gates and candidate in one go
        gates = np.dot(combined, self.W) + self.b

        # Split into 4 parts
        i_gate = sigmoid(gates[:, :self.hidden_size])           # Input gate
        f_gate = sigmoid(gates[:, self.hidden_size:2*self.hidden_size])  # Forget gate
        o_gate = sigmoid(gates[:, 2*self.hidden_size:3*self.hidden_size]) # Output gate
        c_candidate = np.tanh(gates[:, 3*self.hidden_size:])    # Candidate values

        # Update cell state
        c_new = f_gate * c_prev + i_gate * c_candidate
        # The forget gate "forgets" old memories (scales by f_gate)
        # The input gate decides what to remember (scales by i_gate)

        # Compute new hidden state
        h_new = o_gate * np.tanh(c_new)
        # The output gate filters what to output

        return h_new, c_new

def sigmoid(x):
    return 1 / (1 + np.exp(-x))

# Forget gate: f_gate = σ(Wf · [h_{t-1}, x_t] + b_f)
#   → Controls what to forget from previous cell state
# Input gate: i_gate = σ(Wi · [h_{t-1}, x_t] + b_i)
#   → Controls what new information to add
# Candidate: c_candidate = tanh(Wc · [h_{t-1}, x_t] + b_c)
#   → Proposed new cell values
# Output gate: o_gate = σ(Wo · [h_{t-1}, x_t] + b_o)
#   → Controls what to output to next layer/timestep

# Example: Processing "The student studied hard"
# Step 1: see "The" → forget gate ≈ 0, input gate ≈ 1 → new cell state ≈ "The"
# Step 2: see "student" → forget gate ≈ 1, input gate ≈ 0.7 → cell state ≈ "The student"
# Step 3: see "studied" → forget gate ≈ 0.8, input gate ≈ 0.9 → cell state ≈ "student studied"
# Step 4: see "hard" → forget gate ≈ 0.7, input gate ≈ 1 → cell state ≈ "student studied hard"

Why LSTM Works:

• Forget gate: Can learn when to "delete" old memories
• Input gate: Can learn when new information is important
• Additive update: c_new = f_gate × c_prev + i_gate × c_candidate uses addition (not multiplication through tanh). Gradients flow better! (∂c_new/∂c_prev includes f_gate term, which can be close to 1)
• Long-term memory: Cell state is separate from hidden state. Can maintain information across many steps.

Part 5: Applications—Text and Time Series

Text Generation:

# Train on Shakespeare, let it generate text

def generate_text(model, seed="To", length=100):
    """
    model: trained RNN/LSTM
    seed: starting text
    length: characters to generate
    """
    text = seed
    hidden_state = initialize_hidden_state(model.hidden_size)

    for _ in range(length):
        # Get probability distribution over next character
        logits = model.forward(text[-1], hidden_state)
        probabilities = softmax(logits)

        # Sample next character (can also use argmax for deterministic)
        next_char = np.random.choice(vocabulary, p=probabilities)
        text += next_char

    return text

# Output might be:
# "To be or not to be, that is the question:
#  Whether 'tis nobler in the mind to suffer..."

# Pretty good for random initialization!

Stock Price Prediction:

# Time series: past 30 days of stock price → predict next day

def lstm_price_prediction():
    # Input: 30 days of prices
    X = np.array([
        [100, 102, 101, 103, 105, ...],  # Stock prices for 30 days
        [95, 97, 96, 98, 100, ...],
        # ... more sequences
    ])

    # Target: next day's price
    y = np.array([107, 102, ...])

    # LSTM processes all 30 values, learns temporal patterns
    # ("if prices went up last 5 days, likely continues up")
    # ("if volatility high, likely crash imminent")
    # ("seasonal patterns: stock often up on Fridays")

    model = LSTM(input_size=1, hidden_size=64, output_size=1)
    model.train(X, y, epochs=100)

    # Predict tomorrow's price
    tomorrow = model.predict(X[-1:])  # [104.50]

Machine Translation (Encoder-Decoder):

# Translate Hindi to English using RNNs

# Encoder: process Hindi sentence
# "मुझे एक कप कॉफी चाहिए" (I need a cup of coffee)

encoder = LSTM(vocab_size_hindi, hidden_size=256)
final_hidden, final_cell = encoder.process_sequence(hindi_sentence)
# final_hidden is a vector encoding the entire sentence meaning

# Decoder: generate English word by word
decoder = LSTM(vocab_size_english, hidden_size=256)
decoder.initialize_with(final_hidden, final_cell)

english = ""
for _ in range(max_length):
    word_logits = decoder.forward(english[-1])  # Predict next word
    next_word = sample_word(word_logits)
    english += next_word
    if next_word == END_TOKEN:
        break

# Output: "I need a cup of coffee"

Part 6: Bidirectional RNNs and Attention

Problem: RNN only sees past. To understand "The bank of the river", it needs to see "river" comes later.

Bidirectional RNN: Process sequence forwards AND backwards, concatenate hidden states.

# Forward RNN processes: "The" → "bank" → "of" → "the" → "river"
# Backward RNN processes: "river" → "the" → "of" → "bank" → "The"

# When processing "bank":
# - Forward hidden state: context of "The bank"
# - Backward hidden state: context of "bank of the river"
# - Concatenated: full context!

# This is why bidirectional RNNs work better for many tasks (NER, POS tagging, etc.)

Attention Mechanism: Learn which words matter most.

# Translation: "मुझे एक कप कॉफी चाहिए"
# English: "I need a cup of coffee"

# Naive decoder: uses only last encoder state
# Problem: loses information (30-word sentence compressed to one vector)

# Attention: at each decoding step, look back at encoder states
# When generating "coffee", focus on "कॉफी" (coffee)
# When generating "cup", focus on "कप" (cup)

# Attention weights α_ij: how much does output word i depend on input word j?
# Usually: α_ij ∈ [0, 1], Σ_j α_ij = 1

# Context vector = Σ_j α_ij * encoder_hidden_j
# This context goes to decoder instead of just final encoder state

# Result: much better translation quality!

Part 7: Real Applications in India

• Bhashini (by NITI Aayog): Hindi ↔ English translation using transformer (attention-based RNN)
• IVERI: Indian vehicle data analysis using LSTM to predict traffic patterns
• AI tutoring: RNNs analyze student interaction patterns to predict learning gaps
• Agricultural forecasting: LSTM models seasonal patterns in rainfall, crop health

Part 8: Modern Architectures—Transformers (Bonus)

Transformers (used in ChatGPT, BERT) solve LSTM limitations:

• Parallelization: LSTM processes sequences sequentially (slow). Transformers process all tokens at once (fast)
• Long-range dependencies: Attention can directly access any previous token (no vanishing gradients)
• Scaling: Transformers scale to billions of parameters (GPT-3, GPT-4)

You'll study transformers in detail at college if you choose ML. For now, understand that RNNs/LSTMs laid the foundation—transformers build on the same principles (attention, memory, sequence modeling) but with different architecture.

Deep Dive: Recurrent Neural Networks and Sequence Models

At this level, we stop simplifying and start engaging with the real complexity of Recurrent Neural Networks and Sequence Models. In production systems at companies like Flipkart, Razorpay, or Swiggy — all Indian companies processing millions of transactions daily — the concepts in this chapter are not academic exercises. They are engineering decisions that affect system reliability, user experience, and ultimately, business success.

The Indian tech ecosystem is at an inflection point. With initiatives like Digital India and India Stack (Aadhaar, UPI, DigiLocker), the country has built technology infrastructure that is genuinely world-leading. Understanding the technical foundations behind these systems — which is what this chapter covers — positions you to contribute to the next generation of Indian technology innovation.

Whether you are preparing for JEE, GATE, campus placements, or building your own products, the depth of understanding we develop here will serve you well. Let us go beyond surface-level knowledge.

Transformer Architecture: The Engine Behind GPT and Modern AI

The Transformer architecture, introduced in the landmark 2017 paper "Attention Is All You Need," revolutionised NLP and eventually all of deep learning. Here is the core mechanism:

# Self-Attention Mechanism (simplified)
import numpy as np

def self_attention(Q, K, V, d_k):
    """
    Q (Query): What am I looking for?
    K (Key):   What do I contain?
    V (Value): What do I actually provide?
    d_k:       Dimension of keys (for scaling)
    """
    # Step 1: Compute attention scores
    scores = np.matmul(Q, K.T) / np.sqrt(d_k)

    # Step 2: Softmax to get probabilities
    attention_weights = softmax(scores)

    # Step 3: Weighted sum of values
    output = np.matmul(attention_weights, V)
    return output

# Multi-Head Attention: Run multiple attention heads in parallel
# Each head learns different relationships:
# Head 1: syntactic relationships (subject-verb agreement)
# Head 2: semantic relationships (word meanings)
# Head 3: positional relationships (word order)
# Head 4: coreference (pronoun → noun it refers to)

The key insight of self-attention is that every token can attend to every other token simultaneously (unlike RNNs which process sequentially). This parallelism enables efficient GPU training. The computational complexity is O(n²·d) where n is sequence length and d is dimension, which is why context windows are a major engineering challenge.

State-of-the-art developments include: sparse attention (reducing O(n²) to O(n·√n)), mixture of experts (MoE — activating only a subset of parameters per input), retrieval-augmented generation (RAG — grounding responses in external documents), and constitutional AI (alignment through principles rather than RLHF alone). Indian researchers at institutions like IIT Bombay, IISc Bangalore, and Microsoft Research India are actively contributing to these frontiers.

India's Scale Challenges: Engineering for 1.4 Billion

Building technology for India presents unique engineering challenges that make it one of the most interesting markets in the world. UPI handles 10 billion transactions per month — more than all credit card transactions in the US combined. Aadhaar authenticates 100 million identities daily. Jio's network serves 400 million subscribers across 22 telecom circles. Hotstar streamed IPL to 50 million concurrent viewers — a world record. Each of these systems must handle India's diversity: 22 official languages, 28 states with different regulations, massive urban-rural connectivity gaps, and price-sensitive users expecting everything to work on ₹7,000 smartphones over patchy 4G connections. This is why Indian engineers are globally respected — if you can build systems that work in India, they will work anywhere.

Advanced Algorithms: Dynamic Programming and Graph Theory

Dynamic Programming (DP) solves complex problems by breaking them into overlapping subproblems. This is a favourite in competitive programming and interviews:

# Longest Common Subsequence — classic DP problem
# Used in: diff tools, DNA sequence alignment, version control

def lcs(s1, s2):
    m, n = len(s1), len(s2)
    dp = [[0] * (n + 1) for _ in range(m + 1)]

    for i in range(1, m + 1):
        for j in range(1, n + 1):
            if s1[i-1] == s2[j-1]:
                dp[i][j] = dp[i-1][j-1] + 1
            else:
                dp[i][j] = max(dp[i-1][j], dp[i][j-1])

    return dp[m][n]

# Dijkstra's Shortest Path — used by Google Maps!
import heapq

def dijkstra(graph, start):
    dist = {node: float('inf') for node in graph}
    dist[start] = 0
    pq = [(0, start)]  # (distance, node)

    while pq:
        d, u = heapq.heappop(pq)
        if d > dist[u]:
            continue
        for v, weight in graph[u]:
            if dist[u] + weight < dist[v]:
                dist[v] = dist[u] + weight
                heapq.heappush(pq, (dist[v], v))

    return dist

# Real use: Google Maps finding shortest route from
# Connaught Place to India Gate, considering traffic weights

Dijkstra's algorithm is how mapping applications find optimal routes. When you ask Google Maps to navigate from Mumbai to Pune, it models the road network as a weighted graph (intersections are nodes, roads are edges, travel time is weight) and runs a variant of Dijkstra's algorithm. Indian highways, city roads, and even railway networks can all be modelled this way. IRCTC's route optimisation for trains across 13,000+ stations uses graph algorithms at its core.

Research Frontiers and Open Problems in Recurrent Neural Networks and Sequence Models

Beyond production engineering, recurrent neural networks and sequence models connects to active research frontiers where fundamental questions remain open. These are problems where your generation of computer scientists will make breakthroughs.

Quantum computing threatens to upend many of our assumptions. Shor's algorithm can factor large numbers efficiently on a quantum computer, which would break RSA encryption — the foundation of internet security. Post-quantum cryptography is an active research area, with NIST standardising new algorithms (CRYSTALS-Kyber, CRYSTALS-Dilithium) that resist quantum attacks. Indian researchers at IISER, IISc, and TIFR are contributing to both quantum computing hardware and post-quantum cryptographic algorithms.

AI safety and alignment is another frontier with direct connections to recurrent neural networks and sequence models. As AI systems become more capable, ensuring they behave as intended becomes critical. This involves formal verification (mathematically proving system properties), interpretability (understanding WHY a model makes certain decisions), and robustness (ensuring models do not fail catastrophically on edge cases). The Alignment Research Center and organisations like Anthropic are working on these problems, and Indian researchers are increasingly contributing.

Edge computing and the Internet of Things present new challenges: billions of devices with limited compute and connectivity. India's smart city initiatives and agricultural IoT deployments (soil sensors, weather stations, drone imaging) require algorithms that work with intermittent connectivity, limited battery, and constrained memory. This is fundamentally different from cloud computing and requires rethinking many assumptions.

Finally, the ethical dimensions: facial recognition in public spaces (deployed in several Indian cities), algorithmic bias in loan approvals and hiring, deepfakes in political campaigns, and data sovereignty questions about where Indian citizens' data should be stored. These are not just technical problems — they require CS expertise combined with ethics, law, and social science. The best engineers of the future will be those who understand both the technical implementation AND the societal implications. Your study of recurrent neural networks and sequence models is one step on that path.

Key Vocabulary

Here are important terms from this chapter that you should know:

🏗️ Architecture Challenge

Design the backend for India's election results system. Requirements: 10 lakh (1 million) polling booths reporting simultaneously, results must be accurate (no double-counting), real-time aggregation at constituency and state levels, public dashboard handling 100 million concurrent users, and complete audit trail. Consider: How do you ensure exactly-once delivery of results? (idempotency keys) How do you aggregate in real-time? (stream processing with Apache Flink) How do you serve 100M users? (CDN + read replicas + edge computing) How do you prevent tampering? (digital signatures + blockchain audit log) This is the kind of system design problem that separates senior engineers from staff engineers.

The Frontier

You now have a deep understanding of recurrent neural networks and sequence models — deep enough to apply it in production systems, discuss tradeoffs in system design interviews, and build upon it for research or entrepreneurship. But technology never stands still. The concepts in this chapter will evolve: quantum computing may change our assumptions about complexity, new architectures may replace current paradigms, and AI may automate parts of what engineers do today.

What will NOT change is the ability to think clearly about complex systems, to reason about tradeoffs, to learn quickly and adapt. These meta-skills are what truly matter. India's position in global technology is only growing stronger — from the India Stack to ISRO to the startup ecosystem to open-source contributions. You are part of this story. What you build next is up to you.

Crafted for Class 10–12 • Deep Learning & NLP • Aligned with NEP 2020 & CBSE Curriculum