Frequently Asked Questions About Edureka AI & ML Full Course Skill
22 answers covering everything from basics to advanced usage.
// Basics
What does 'AI is a broader umbrella' actually mean?
It means artificial intelligence, machine learning, and deep learning form a nested hierarchy, not a set of synonyms. AI is the outermost concept — any technique that makes machines intelligent. Machine learning is the subset that learns from data. Deep learning is the further subset using neural networks. Treating them as interchangeable leads to wrong tool selection.
What are the three stages of AI?
Artificial Narrow Intelligence (ANI/Weak AI) performs specific predefined tasks and is where all deployed systems sit today. Artificial General Intelligence (AGI/Strong AI) would let machines think and reason like humans — not yet achieved. Artificial Super Intelligence (ASI) would surpass human capability in all domains and currently exists only in science fiction.
What are the four types of AI by functionality?
Reactive Machines use only present data with no memory, like Deep Blue. Limited Memory AI uses recent past data to decide, like self-driving cars reading sensors. Theory of Mind AI focuses on emotional intelligence and isn't fully developed. Self-Aware AI, with machine consciousness, is hypothetical and corresponds to Artificial Super Intelligence.
What is the target variable versus a predictor variable?
The target variable is the output your model is built to predict — for example, whether a shipment is delayed. Predictor variables are the input features the model uses to make that prediction, like route, weather, and carrier. Supervised learning requires a target variable; unsupervised learning has none because it discovers patterns instead.
// How To
How do I define the objective for my ML project?
Start by clarifying exactly what needs to be predicted or discovered. Ask: what is the target variable? Is this regression, classification, or clustering? What input features are available? Answering these first prevents scope drift before you touch any data and determines every subsequent decision, especially algorithm choice.
How do I split my data into training and testing sets?
Use data splicing to divide your dataset before building the model — one portion trains the model, the other tests it on unseen data. Never evaluate on the same data you trained on, as that produces falsely optimistic accuracy. During evaluation, apply cross-validation, one of the easiest and most reliable ways to check real accuracy.
How do I handle an NLP problem like analyzing reviews?
Treat it as a natural language processing task using the NLTK library. Apply tokenization to break text into units, then stemming and lemmatization to normalize words. If you have labeled sentiment data, frame it as supervised classification; if not, use an unsupervised clustering approach to group sentiment. Then interpret which reviews are positive, negative, or neutral.
How do I choose an algorithm once I know my problem type?
Map the problem type to the algorithm family. For regression use linear regression; for classification use logistic regression, decision trees, random forest, SVM, or KNN; for clustering use K-Means; for reward-based sequential problems use Q-Learning. Implement with the matching Python library — Scikit-learn, TensorFlow, Keras, or NLTK depending on the task.
How do I decide between classification and regression?
Look at your desired output type. If you need a continuous numeric prediction, like price or temperature, it's regression. If you need a category, like spam/not-spam or delayed/on-time, it's classification. Both fall under supervised learning, but the distinction determines your algorithm choice and how you measure accuracy.
// Troubleshooting
My model has high accuracy on training data but fails on new data — what happened?
You likely evaluated on the same data you trained on, or skipped proper data splicing. Split your dataset into separate training and testing sets, then use cross-validation to get honest accuracy. Falsely optimistic training accuracy is a classic sign of overfitting — the model memorized rather than generalized.
My deep learning model performs poorly — could my dataset be the problem?
Yes. Deep learning is like a rocket engine that requires massive data to function well. On small datasets it underperforms and wastes GPU-heavy infrastructure. Switch to a traditional ML algorithm like decision trees, logistic regression, or random forest, which are more appropriate when data volume is low.
My predictions are consistently wrong — where should I look first?
Look at your data pre-processing. Wrongful computation downstream is almost always caused by dirty data — missing values, duplicates, or redundant variables left uncleaned. Re-scan the entire dataset, impute or remove inconsistencies, then rerun EDA to confirm your selected features actually carry predictive power before rebuilding.
Regulators want to know why my model made a decision, but I can't explain it. What now?
You're facing the black box problem. Deep learning models are mathematically traceable at the node level but not interpretable in human terms. In regulated or high-stakes environments, switch to interpretable models like decision trees or logistic regression, which produce crisp, explainable rules. Document what your model can and cannot explain.
// Comparisons
How does supervised learning compare to unsupervised learning?
Supervised learning trains on labeled data under external supervision and solves regression and classification problems where the output is known. Unsupervised learning trains on unlabeled data with no guidance, autonomously discovering patterns and forming clusters. Choose supervised when you have labels and a known target; choose unsupervised when the goal is pattern discovery or grouping.
How does feature engineering differ between ML and deep learning?
In traditional machine learning, a domain expert must identify and handcode features manually to reduce complexity and expose patterns. Deep learning automatically learns features from raw data — first low-level ones like edges and lines, then increasingly abstract representations — removing manual feature extraction. The tradeoff is deep learning's need for far more data.
How does this framework compare to jumping straight into a Kaggle notebook?
A Kaggle notebook usually starts mid-process with clean data and a defined problem. This framework starts earlier, forcing you to define the objective, identify the AI/ML type, and select the learning type before coding. That upstream rigor prevents choosing the wrong algorithm family and catches data issues that pre-packaged notebooks hide.
When should I use reinforcement learning instead of supervised learning?
Use reinforcement learning when an agent must act in an environment and optimize for cumulative rewards over time, rather than predict from a labeled dataset. Sequential decision problems like game strategy or robotics fit this. Supervised learning suits static prediction from labeled examples. If there's no target variable but there are rewards and penalties, reinforcement learning applies.
// Advanced
What is cross-validation and how does it improve my model?
Cross-validation is one of the most important and easiest methods for checking model accuracy during evaluation. Instead of a single train/test split, it repeatedly partitions the data to test the model on multiple subsets, giving a more reliable accuracy estimate. Combined with parameter tuning, it helps you decide whether accuracy is acceptable or needs more iteration.
How does the Turing Test relate to modern AI?
The Turing Test, proposed by Alan Turing in 1950, judges machine intelligence: if a human evaluator can't distinguish a machine from a human through text conversation, the machine passes. It's a conceptual benchmark for general intelligence (AGI), not a practical evaluation metric for the ANI systems you build day to day, which are measured by accuracy and cross-validation.
How do I know if my problem needs ANI, and which functionality type applies?
Almost all practical systems are ANI performing specific task automation. Within ANI, determine the functionality type: Reactive Machine for present-data-only decisions, or Limited Memory for decisions using recent past data, which covers most real applications. Flag whether the problem is sequential and reward-based, indicating reinforcement learning versus standard supervised or unsupervised approaches.
Can I combine multiple learning types in one project?
Yes, complex projects often blend approaches. For instance, an NLP sentiment task might use unsupervised clustering to explore unlabeled reviews, then supervised classification once labels exist. What matters is matching each sub-problem to the correct learning type — regression and classification to supervised, grouping to unsupervised, reward-based sequences to reinforcement — rather than forcing one method across the whole pipeline.
What Python libraries should I use for each task type?
Use Scikit-learn for most classical ML algorithms like logistic regression, random forest, SVM, KNN, and K-Means. Use TensorFlow and Keras for deep learning and neural networks. Use NLTK for natural language processing tasks like tokenization and lemmatization. NumPy and Theano support numerical computation underneath these. Match the library to your problem type and algorithm.