25 In-Depth AI Terms Every Serious UPSC Aspirant Should Master

I. ⚙️ ARCHITECTURE & CORE WORKINGS

1. Backpropagation
   A training technique where the model adjusts its internal weights by calculating the error and minimizing it—like correcting itself after a mistake.
2. Weights and Biases
   Parameters inside neural networks that get updated during training to learn patterns in data. Like memory markers of what matters.
3. Epoch
   One full cycle where the algorithm sees the entire dataset once during training. Multiple epochs = better learning.
4. Loss Function
   A metric that tells the model how wrong it is. The aim is to minimize this value over training iterations.
5. Gradient Descent
   Optimization algorithm used to find the minimum loss. Like taking the steepest downhill path to the answer.

---

II. 🧠 FUNCTIONAL INTELLIGENCE

6. Attention Mechanism
   Allows AI to focus on relevant parts of the input. In translation, it helps the model understand “what to pay attention to.”
7. Self-Attention
   A variant of attention where each word in a sentence attends to every other word—key in transformers like GPT and BERT.
8. Transfer Learning
   Pretrained models on one task reused for another. Saves compute and time. Used in Indic language models under the IndiaAI Mission.
9. Zero-shot / Few-shot Learning
   The model answers questions it hasn’t explicitly been trained for, using general knowledge.
10. Fine-Tuning
    Adjusting a pre-trained model on specific, domain-based data (like Indian languages or medical data).

---

III. 💻 DATA, MODEL TYPES, TRAINING STYLES

11. Overfitting vs Underfitting

• Overfitting: Model is too tailored to training data, fails on new inputs
• Underfitting: Model is too simple, misses the patterns

12. Dropout
    A regularization technique where some neurons are turned off during training to prevent overfitting.
13. Batch Size
    The number of training examples used in one iteration. Impacts memory use and speed.
14. Hyperparameters
    Configurable settings (like learning rate, batch size) set before training. Optimizing these = better model.
15. Cross-validation
    A way to test the model’s generalization by splitting data into multiple parts. Ensures reliability.

---

IV. 📡 AI APPLICATION ENGINES

16. Embeddings
    Numerical representations of data (like words) in vector form. Helps machines “understand” relationships between words.
17. Word2Vec / Doc2Vec
    Embedding techniques where the AI learns semantic meanings from context.

Foundation for sentiment analysis tools in governance.

18. Latent Space
    The compressed space in which models organize their understanding of the world.

Used in image generation models like Stable Diffusion.

19. Autoencoders
    Neural networks that compress input into low dimensions and reconstruct it back.

Useful for anomaly detection in fraud analytics.

20. Generative Adversarial Networks (GANs)
    Two neural networks—Generator and Discriminator—compete to improve realism.

GANs are behind deepfakes, synthetic image generation.

---

V. 🛡️ ETHICS, SAFETY & POLICY

21. Hallucination (AI)
    When GenAI makes up facts that are incorrect or unverifiable.

A major challenge for AI in governance & education.

22. Data Drift
    When the input data over time changes and the model starts performing poorly.

Important in real-time systems like traffic prediction.

23. Model Explainability
    How well humans can interpret the reasoning behind AI decisions.

Key for AI use in judiciary, healthcare, and policing.

24. Bias Amplification
    When AI not only reflects but amplifies existing social biases in the training data.

Seen in loan approvals, hiring, predictive policing.

25. Federated Learning
    A privacy-focused technique where models train across multiple devices without sharing raw data.

Relevant to Data Governance & PDP Act 2023 discussions.

---

🧠 BONUS: How These Terms Connect in Real World

Let’s say you want to build a chatbot in an Indian language:

• You’d use a pre-trained LLM and fine-tune it on Bhashini datasets
• Employ transfer learning and embeddings for language understanding
• Apply attention mechanism to handle complex conversations
• Ensure model explainability for transparency
• Use federated learning for privacy if deployed in sensitive sectors like health or finance