Frequently Asked Questions About Edureka AI/ML Foundations Builder
22 answers covering everything from basics to advanced usage.
// Basics
What does 'AI is an umbrella' actually mean in practice?
It means AI, machine learning, and deep learning are nested, not parallel. AI is the broadest goal of making machines intelligent. Machine learning is one data-driven way to achieve AI. Deep learning is a neural-network-based subset of ML. Data science is the widest umbrella and uses all three. In practice, you never 'choose AI vs ML' — you choose a technique within the hierarchy.
What is Artificial Narrow Intelligence and are all current AI systems narrow?
Artificial Narrow Intelligence, also called Weak AI, is the current stage where machines perform only a narrowly defined set of specific tasks with no general reasoning. Yes — every commercially deployed AI system today, including Siri, Alexa, AlphaGo, and self-driving cars, is narrow. Artificial General Intelligence (human-level reasoning) and Artificial Super Intelligence remain future or hypothetical stages.
What is data splicing and why can't I skip it?
Data splicing is dividing your input data into a training data set (used to build the model) and a testing data set (used to evaluate it). The training set is always larger. You can't skip it because evaluating a model on the same data it learned from produces misleadingly high accuracy — the model has effectively seen the answers. Always split before any evaluation.
Why is Python the standard language for AI and machine learning?
Python dominates because it requires less coding via a check-as-you-code methodology, has pre-built libraries covering every algorithm, uses simple English-like syntax that's easy to learn, offers platform independence through tools like PyInstaller, and has massive community support. Critically, all major AI/ML libraries — Scikit-learn, TensorFlow, Keras, NumPy, NLTK — are Python-native.
What is the Turing Test and does passing it prove intelligence?
The Turing Test, proposed by Alan Turing in 1950, states that if a human evaluator cannot distinguish a machine from a human through text-based communication, the machine is said to have passed. It's a benchmark for conversational indistinguishability rather than proof of genuine understanding or consciousness. It remains a foundational concept when discussing what machine intelligence means.
// How To
How do I run the seven-step process for a fraud detection project?
Define the objective (predict fraud/not-fraud, a classification problem). Gather labeled transaction records. In data preparation, handle class imbalance since fraud is rare. Run EDA to find fraud-correlated features. Split into train/test, then train Logistic Regression for interpretability. Evaluate with cross-validation. Deploy for predictions. Use Scikit-learn throughout, and favor Logistic Regression or Decision Trees if auditors require explanations.
How do I preprocess text data before building a classifier?
Use NLTK for text pre-processing: tokenization, stemming, lemmatization, and stopword removal. For an email routing classifier, remove stopwords and apply stemming before feature extraction, then feed the cleaned features into a Scikit-learn model like Naive Bayes, a classic baseline for text classification. This combines the NLP and Machine Learning branches within a single supervised classification workflow.
How do I decide whether to remove a variable during data preparation?
Check two things separately. First, remove variables with excessive null rates — over 40% is a common threshold. Second, remove variables that leak information about the target, such as a 'risk_mm' field that directly reveals the outcome. Both must be removed but for different reasons: noisy variables add nothing, while leaking variables inflate accuracy artificially and corrupt the model silently.
How do I set up reinforcement learning for a warehouse navigation robot?
Use reinforcement learning because there's no predefined data — the agent learns by trial and error. Place the agent in the warehouse environment, give rewards for correct navigation and penalties for collisions. Start with Q-Learning as a baseline. The agent explores, collects its own data, and iteratively improves its policy. If the state space is too large for classical Q-Learning, layer a Deep Q-Network on top.
How much time should I budget for data preparation in an ML project?
Budget the majority of your project time for data preparation — it is explicitly the most time-consuming step in the seven-step process. You'll scan for missing values, null values, duplicates, and redundant variables, remove high-null-rate columns, eliminate leaking variables, and handle outliers. Rushing this step corrupts the model silently, so treat it as the foundation, not a formality.
// Troubleshooting
Why does my deep learning model perform worse than a simple algorithm?
Most likely your data set is too small. Deep learning needs large amounts of data to learn patterns perfectly; on small data sets, classical ML algorithms outperform it. Match algorithm class to data volume before choosing a specific algorithm. If your data is limited, switch to Scikit-learn models like Random Forest or Logistic Regression rather than forcing a neural network.
My model shows high accuracy but fails in production — what went wrong?
You likely evaluated on training data or had a leaking variable. If you skipped data splicing, accuracy reflects memorization, not generalization. If a predictor variable secretly revealed the target, accuracy was inflated during training but collapses on genuinely unseen data. Re-split with a proper training/testing set and audit your features for information leakage.
Why is my deep learning training taking so long?
Deep learning trains slowly by nature — potentially weeks — because it performs heavy matrix multiplication across many layers. If you're on a CPU-only machine, that's your bottleneck: deep learning requires GPUs, which aren't optional. Either provision GPU compute (via TensorFlow's GPU support) or reconsider whether classical ML on Scikit-learn meets your needs, since it trains far faster.
I picked a neural network but auditors reject it — what should I do?
Switch to an interpretable model. Neural networks are black boxes: you can identify which nodes activated but not what they collectively represent, so you can't explain decisions to auditors. In regulated use cases, replace the neural network with a Decision Tree or Logistic Regression, which provide crisp, auditable rules. Trade some accuracy for the explainability your regulatory environment demands.
// Comparisons
How does machine learning feature engineering compare to deep learning?
In machine learning, a domain expert must identify and hand-code features. In deep learning, the algorithm learns high-level features automatically — for example, CNNs progress from edges to face parts to a full face representation. Choose machine learning when domain knowledge is rich and data is scarce; choose deep learning when data is abundant and manual feature extraction is impractical.
How does end-to-end deep learning compare to decomposed ML problem solving?
Machine learning typically breaks a problem into sub-parts, solves each, then combines results. Deep learning solves problems end-to-end — YOLO Net takes an image and returns both the object name and location in a single step. Choose the approach that matches your complexity and latency needs: decomposition offers control and interpretability, while end-to-end offers simplicity and speed at inference.
How does Scikit-learn compare to TensorFlow and Keras?
Scikit-learn is for general classical ML — classification, regression, clustering, and dimensionality reduction. TensorFlow is Google-backed for production deep learning with CPU+GPU parallel neural network training. Keras is for quick neural network prototyping, running on CPU+GPU with Python-native debuggability and support for all major model types. Use Scikit-learn for classical ML, Keras to prototype, and TensorFlow to ship deep learning.
How does supervised learning compare to unsupervised learning in output?
Supervised learning is trained on labeled data acting as a teacher, and solves regression (continuous output) and classification (categorical output) problems. Unsupervised learning is trained on unlabeled data with no guidance, discovering patterns on its own, and solves clustering (grouping by similarity) and association analysis (co-occurrence rules) problems. The presence or absence of labels is the deciding factor.
// Advanced
How do KNN and other classical algorithms behave at test time as data grows?
Classical ML trains fast but some algorithms slow down at test time as data grows — K-Nearest Neighbour is the classic example, since it compares each query against stored points. Deep learning is the opposite: it trains slowly, potentially over weeks, but tests fast once trained. Factor test-time latency into your algorithm choice, not just training time.
When should I use the Apriori algorithm over clustering?
Use Apriori for association analysis — finding co-occurrence rules, like market basket analysis showing which products are bought together. Use K-Means clustering when you want to group data points by feature similarity, such as segmenting users by viewing habits. Both are unsupervised, but Apriori finds rules between items while clustering finds groups of similar records.
Can a single problem span multiple AI branches?
Yes. A problem may span Machine Learning, Deep Learning, NLP, Computer Vision, Robotics, Expert Systems, or Fuzzy Logic simultaneously. For example, a customer support email router combines NLP for text pre-processing with Machine Learning for classification. A warehouse robot combines Reinforcement Learning with Robotics. Map to all relevant branches rather than forcing your problem into one.
What is the deep learning decision filter and when do I apply it?
Apply the deep learning decision filter only after classical ML performance proves insufficient. Switch to deep learning only if all four hold: data volume is large enough, GPU compute is available, end-to-end solving beats decomposition, and interpretability is not a hard requirement. If any condition fails, stay with classical ML — deep learning's cost and black-box nature won't pay off.