AHSEC Class 12 Artificial Intelligence (Subject 527): Comprehensive Marks Weightage, Blueprint & 50-Mark Theory Master Guide

1. Executive Overview: The Higher Secondary AI Framework in Assam

With the notification of the National Education Policy (NEP 2020) and the restructuring of the Assam Higher Secondary Education Council (AHSEC) into the Assam State School Education Board (ASSEB), Artificial Intelligence was officially inducted into senior secondary education across the state.

Subject Code 527 is not restricted to Science stream learners. Students enrolled in Arts, Science, and Commerce are equally eligible to choose Artificial Intelligence as an elective subject. The curriculum is structurally aligned with the Central Board of Secondary Education (CBSE Subject Code 843) and national NSQF Level 4 occupational standards, but incorporates indigenous North-Eastern ecological and economic case studies—such as tea plantation pest telemetry, Brahmaputra hydrological monitoring, and indigenous handloom pattern synthesis.

2. The 100-Mark Dual Examination Structure

Unlike conventional Science electives (Physics, Chemistry, Biology) which follow a 70-mark theory and 30-mark practical division, Artificial Intelligence adopts an equally weighted 50–50 dual structure.

3. Complete Chapter-Wise Marks Weightage Blueprint

The 50-mark theory paper is strictly divided into two segments: Part A (Employability Skills — 10 Marks) and Part B (Subject-Specific AI Skills — 40 Marks).

Part A: Employability Skills (10 Marks Weightage)

Prescribed by the PSSCIVE (Pandit Sunderlal Sharma Central Institute of Vocational Education), Bhopal, Part A evaluates workplace readiness and environmental consciousness.

Part B: Subject-Specific AI Competencies (40 Marks Weightage)

Part B forms the core technical syllabus for Class 12, carrying 80% of the written exam weightage.

4. Deep Syllabus Breakdown: Part A Employability Skills

Unit 1: Communication Skills-IV (4 Marks)

Students must master active listening techniques, barrier minimization in technical teamwork, and formal business communication.

• Key Topics: Non-verbal communication nuances, constructing persuasive technical emails, organizing executive summaries for technical documentation, and delivering 3-minute project pitches.
• Typical Question Pattern: Identifying communication barriers in a multi-disciplinary software engineering team or drafting a professional email requesting API server credentials.

Unit 2: Entrepreneurial Skills-IV (4 Marks)

Focuses on transforming AI competencies into viable socio-economic solutions in Assam and Northeast India.

• Key Topics: Entrepreneurial mindset, business plan canvas, local market problem validation (e.g., automated tea quality grading, Brahmaputra flood alert networks), cost-benefit analysis, and risk mitigation.
• Typical Question Pattern: Formulating a value proposition statement or calculating the break-even point for a localized agritech startup.

Unit 3: Green Skills-IV (2 Marks)

Focuses on sustainable development goals (SDGs) and mitigating the carbon footprint of compute infrastructure.

• Key Topics: Energy consumption of large machine learning models, green data centers, electronic waste (e-waste) disposal legislation in India, and lifecycle assessment of computer hardware.

5. Deep Syllabus Breakdown: Part B Subject-Specific Competencies

Unit 1: Capstone Project & AI Lifecycle (10 Marks)

This unit tests whether the student understands how real-world machine learning systems are conceived, engineered, and audited.

• Problem Scoping: The 4Ws Problem Canvas (Who is affected? What is the nature of the issue? Where does it arise? Why is a machine learning solution advantageous?).
• Data Exploration: Structured tabular data vs unstructured sensor streams; handling missing values (imputation via mean/median vs deletion).
• Evaluation Metrics:
• Accuracy = (TP + TN) / (TP + TN + FP + FN)
• Precision = TP / (TP + FP)
• Recall = TP / (TP + FN)
• F1-Score = 2 (Precision Recall) / (Precision + Recall)
• Board Trap: Examiners frequently ask: "In medical tumor diagnosis or flood warnings, which metric is more critical—Precision or Recall?" (Answer: Recall, because minimizing False Negatives is paramount to prevent life-threatening oversights).

Unit 2: Computer Vision in Practice (10 Marks)

Understanding how computers interpret multidimensional digital images.

• Image Representation: 2D matrices (Grayscale: 0 to 255) and 3D tensors (RGB: Height x Width x 3 Channels).
• Convolutional Filtering: Kernel math (3x3 matrices sliding across pixel grids) for edge detection, Gaussian blurring, and sharpening.
• OpenCV Fundamentals: cv2.imread(), cv2.cvtColor(img, cv2.COLOR_BGR2GRAY), cv2.Canny(), and morphological transformations (Dilation & Erosion).

Unit 3: Natural Language Processing (10 Marks)

Transforming raw, unstructured vernacular and English text into numerical vectors that neural networks can process.

• Preprocessing Pipeline: Sentence segmentation -> Word tokenization -> Lowercasing -> Stopword removal -> Stemming (heuristic truncation like Porter Stemmer) vs Lemmatization (vocabulary-informed morphological reduction like WordNet).
• Vectorization: Bag-of-Words (BoW) frequency counts vs Term Frequency-Inverse Document Frequency (TF-IDF):
• TF(t, d) = Count of t in d / Total words in d
• IDF(t, D) = log(Total documents N / Documents containing t)
• Applications: Sentiment analysis of customer reviews and automated categorization of regional news articles.

Unit 4: Neural Networks & Deep Learning (10 Marks)

Moving from classical statistical estimators to multi-layer perceptron (MLP) architectures.

• The Perceptron:
• Weighted Sum: z = Sum(wi * xi) + b
• Output: y_hat = f(z)
• Activation Functions:
• Sigmoid: 1 / (1 + e^(-z)) (Bound between 0 and 1; used in binary output layers).
• ReLU (Rectified Linear Unit): max(0, z) (Solves vanishing gradient in hidden layers).
• Softmax: Multi-class probability distribution normalization.
• Optimization: Understanding how gradient descent adjusts weights by backpropagating loss calculated via mean squared error or cross-entropy.

6. Part C: 50-Mark Board Practical Exam Rubric

The practical examination is conducted at the student's institutional computer lab and evaluated by an external examiner appointed by ASSEB.

Top 5 Viva Questions Asked by External Examiners (With Model Answers)

1. 1.
   Question: "What is the difference between supervised and unsupervised learning?"

• Answer: Supervised learning trains models on labeled datasets with explicit ground-truth targets (e.g., classification, regression). Unsupervised learning uncovers latent patterns, clusters, or dimensionality reductions in unlabeled data without prior targets (e.g., K-Means, PCA).

1. 1.
   Question: "Why do we use ReLU instead of Sigmoid in deep hidden layers?"

• Answer: Sigmoid saturates at both extremes (outputs close to 0 or 1), causing its derivative to approach zero, which triggers the 'vanishing gradient problem' and stalls backpropagation. ReLU maintains a constant gradient of 1 for all positive inputs, ensuring faster convergence.

1. 1.
   Question: "What happens if you train a model without shuffling your training dataset?"

• Answer: The model may learn spurious chronological order or class clustering biases, leading to poor generalization and high validation error.

1. 1.
   Question: "Explain the difference between stemming and lemmatization."

• Answer: Stemming applies heuristic rule-based cutting of word suffixes (often generating non-words like 'carri' from 'caring'), whereas lemmatization uses lexical databases to resolve words to their legitimate grammatical root (lemma) based on part-of-speech context.

1. 1.
   Question: "What is a confusion matrix?"

• Answer: A 2x2 performance measurement table showing the four combinations of predicted vs actual outcomes: True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN).

7. Model Questions & Authentic Scoring Keys

1-Mark Multiple Choice Questions

Q1: In natural language processing, which technique reduces words to their meaningful dictionary base form based on vocabulary and morphological analysis?

• (A) Stemming
• (B) Lemmatization
• (C) Tokenization
• (D) Stopword Removal
• Correct Answer: (B) Lemmatization. (Stemming chops off characters without dictionary validation).

Q2: A model classifying tea leaves detects 90 out of 100 diseased leaves correctly. However, it mistakenly flags 30 healthy leaves as diseased. What is the Recall of this model?

• (A) 90 / (90 + 30) = 75%
• (B) 90 / 100 = 90%
• (C) 30 / 100 = 30%
• (D) 100 / 90 = 111%
• Correct Answer: (B) 90%. (Recall is TP / Actual Positives = 90 / 100 = 0.90).

2-Mark Short Conceptual Question

Q3: State the purpose of the 4Ws Canvas in the AI Project Cycle.

• Model Answer: The 4Ws Problem Canvas is a structured framework used during the Problem Scoping stage to delineate project boundaries:

1. a.
   Who: Identifies stakeholders facing the issue and the direct beneficiaries of the solution.
2. b.
   What: Defines the nature of the bottleneck and evidence demonstrating its severity.
3. c.
   Where: Outlines the operational environment and context where the problem occurs.
4. d.
   Why: Articulates the tangible socio-economic value generated by resolving the problem through AI.

5-Mark Comprehensive Case Study Question

Q4: A high school in Assam is developing an automated attendance system using facial recognition. (a) Identify the three primary stages of the Computer Vision pipeline required. (b) Explain how illumination differences could impact model accuracy. (c) Suggest two ethical safeguards the school must implement.

• Model Answer:
• (a) Three Pipeline Stages (2 Marks):

1. a.
   Image Acquisition & Preprocessing: Capturing frame from IP camera, converting BGR to Grayscale, and resizing to standardized dimensions (224 x 224).
2. b.
   Feature Extraction: Using Haar cascades or MTCNN to locate face bounding boxes and extracting 128-dimensional facial embedding vectors.
3. c.
   Classification / Matching: Computing cosine similarity or Euclidean distance against enrolled student embedding databases to register attendance.

• (b) Illumination Impact (1.5 Marks): Fluctuations in natural ambient light between morning and overcast monsoon afternoons alter raw pixel intensity matrices drastically, causing facial shadow occlusions. If unmitigated by histogram equalization or adaptive contrast normalization (CLAHE), false rejection rates will escalate.
• (c) Ethical Safeguards (1.5 Marks):

1. a.
   Data Sovereignty & Local Storage: Store biometric embeddings locally in encrypted form on institutional servers without uploading student facial data to public commercial cloud endpoints.
2. b.
   Explicit Parental Consent & Manual Fallback: Comply with India's Digital Personal Data Protection (DPDP) Act by obtaining guardian authorization and maintaining traditional manual attendance registers for opting-out students without penalty.

8. Four Critical Textbook Python Errata Fixes

The official state curriculum printouts contain several typographical bugs that cause Python scripts to throw syntax and execution errors. Mentora has verified the following corrections:

Bug 1: Incorrect OpenCV Display Loop

# ❌ INCORRECT (From Printed Lab Sheet - Causes Window to Freeze)
import cv2
img = cv2.imread('sample.jpg')
cv2.imshow('Display', img)
cv2.destroyAllWindows()  # Throws instant crash without waitKey

# ✅ VERIFIED CORRECTION (Mentora Errata Fix)
import cv2
img = cv2.imread('sample.jpg')
if img is not None:
    cv2.imshow('Display', img)
    cv2.waitKey(0)  # Halts execution until any keyboard key is pressed
    cv2.destroyAllWindows()
else:
    print("Error: Image file not found.")

Bug 2: Missing Random State in Scikit-Learn Split

# ❌ INCORRECT (Inconsistent Evaluation Accuracy Every Run)
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

# ✅ VERIFIED CORRECTION
from sklearn.model_selection import train_test_split
# Pin random_state to ensure reproducible evaluation benchmarks for external examiners
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)

9. 3-Hour Exam Hall Strategy & Time Allocation

To maximize scores on the 50-mark board theory paper, adopt this calibrated 180-minute pacing protocol:

• 0:00 – 0:15 (15 Mins — Cool-Off Reading Period): Carefully inspect Section A and Section B. Identify internal choices in 5-mark long questions where you have verified algorithmic clarity.
• 0:15 – 0:45 (30 Mins — Section A: Employability Skills): Answer all 10 marks of Part A crisply. Avoid essay-style fluff; focus on concise, bulleted business and green skills frameworks.
• 0:45 – 1:30 (45 Mins — Section B: Short 1-Mark & 2-Mark Questions): Complete all MCQs and definitions. Write formula notations explicitly for evaluation metrics.
• 1:30 – 2:35 (65 Mins — Section B: Long Analytical & Case Studies): Draw labeled diagrams for neural network architectures, confusion matrices, and computer vision pipelines. Visual illustrations earn maximum marks from evaluators.
• 2:35 – 3:00 (25 Mins — Revision & Verification): Double-check question numbering, units, formula substitutions, and metric calculations.

10. Official Curriculum Resources & Download Links

Mentora provides free, full-spectrum learning assets for AHSEC Class 11 and 12 Artificial Intelligence:

• [Official AHSEC AI Syllabus (Subject Code 527)](/asseb-ai/syllabus): Complete 100-mark distribution, periods schedule, and PDF download.
• [AHSEC AI Model Question Papers & Scoring Rubrics](/asseb-ai/sample-papers): Authentic solved 50-mark theory paper and viva voce questions.
• [Assam AI Lab-in-a-Box Hub](/asseb-ai): Runnable Python lab files tested for low-spec dual-core computer labs across Assam.