Frequently Asked Questions About Edureka AI/ML Foundations Skill
21 answers covering everything from basics to advanced usage.
// Basics
What is Artificial Narrow Intelligence and why does it matter?
Artificial Narrow Intelligence (ANI/Weak AI) is the first and only currently achieved stage of AI, where machines perform a narrowly defined task with no genuine self-awareness or generalised thinking. Examples include Siri, Alexa, AlphaGo, and self-driving cars. It matters because every deployable production system is ANI — claiming AGI or ASI capability is technically incorrect and misleads stakeholders.
What are the four functional types of AI?
The four functional types are Reactive Machines (present data only, no memory — e.g., IBM Deep Blue), Limited Memory AI (uses recent historical data — e.g., self-driving cars), Theory of Mind AI (emotional and belief comprehension, still in research), and Self-Aware AI (possesses consciousness, hypothetical). Identify which type your system is to set correct expectations.
What is the Turing Test?
The Turing Test, proposed by Alan Turing in 1950, is a benchmark where a human evaluator communicates via text with both a human and a machine. If the evaluator cannot distinguish between them, the machine is said to have passed. It was the first serious proposal in the philosophy of AI and remains a reference point for evaluating machine intelligence.
What is data splicing and what ratio should I use?
Data splicing divides your dataset into a training set (used to build the model, always larger, commonly 70–80%) and a testing set (used only for evaluation, 20–30%). Use Scikit-Learn's train_test_split to apply it. The exact ratio depends on dataset size, but the training set must always be larger and the test set must never be used during training.
// How To
How do I define the objective for a machine learning project precisely?
State three things clearly: what is being predicted (the Target Variable), whether that target is categorical or continuous, and what success looks like. Ambiguity at this stage propagates errors through every downstream step. For example, 'predict shipment delay' should be pinned down as a categorical yes/no target with a defined accuracy or business threshold for success.
How do I prepare and clean my data before modeling?
Scan for and handle missing values, duplicate rows, redundant variables, and incorrectly typed fields using Pandas and NumPy. This is the most time-consuming and most neglected step. Dirty data causes wrongful computation downstream, so never rush it. Load your data into a Pandas DataFrame and immediately record the number of observations and features.
How do I perform Exploratory Data Analysis effectively?
Treat EDA as the brainstorming stage: identify patterns, trends, and correlations between features and the target variable, and map the strong predictors. Visualise distributions and relationships, and flag any class imbalance for classification problems. Insights discovered here directly inform model design — for clustering, EDA also helps you understand what natural groupings exist before choosing K.
How do I build a model using Scikit-Learn?
Select the algorithm identified during problem classification, import it from Scikit-Learn (or TensorFlow/Keras for deep learning, NLTK for NLP), then fit it on the training set only. For deep learning, confirm GPU availability and expect significantly longer training time — potentially weeks from scratch. Always split data first with train_test_split before fitting.
Can I use this framework to explain AI concepts, not just build models?
Yes — one of its three core uses is explaining or classifying AI/ML concepts. The principles distinguish AI, ML, and deep learning, locate systems across the three evolutionary stages and four functional types, and clarify paradigms. This makes it useful for teaching, technical documentation, or setting realistic capability expectations with non-technical stakeholders.
// Troubleshooting
My deep learning model is underperforming on a small dataset — what went wrong?
Deep learning requires large data volumes to outperform classical ML; on small datasets it consistently underperforms. Switch to a classical algorithm like Decision Trees, Random Forest, or Logistic Regression, which win on small data and run on low-end machines. Deep learning's automatic feature learning advantage only pays off when you feed it enough raw data.
My model shows suspiciously high accuracy — what should I check?
Check whether you trained on the full dataset instead of splitting it. Training on test data produces falsely optimistic accuracy because the model was evaluated on data it already saw. Apply data splicing before model building, reserve the testing set (20–30%) for evaluation only, and use cross-validation to confirm the accuracy is genuine.
My cybersecurity model produces too many false positive alerts — why?
Models trained on incomplete datasets commonly produce false positives, leading to alert fatigue and reduced operational efficiency. Establish a baseline of normal behaviour with EDA before training, and be aware that adversarial inputs can exploit model vulnerabilities. Document these risks explicitly, and treat anomaly detection as unsupervised learning since threat labels usually don't exist.
Stakeholders can't understand why my model made a decision — how do I fix that?
You likely deployed a black-box deep learning model where interpretability was required. Replace it with a Decision Tree or Logistic Regression, which provide crisp, inspectable decision rules. In regulated or high-stakes domains, accept a performance cost for explainability. If you must keep a neural network, document the black-box limitation explicitly for stakeholders.
// Comparisons
How does classical machine learning compare to deep learning for feature engineering?
In classical ML, domain experts must manually identify and hand-code features before training. In deep learning, the algorithm automatically learns high-level features from raw data — this is deep learning's most distinctive advantage. Manually engineering features for a deep learning model wastes effort and misuses the technology, so match your effort to the paradigm.
How does the classical ML problem-solving approach differ from deep learning?
Classical ML decomposes problems into sub-parts — for example, object detection first, then object recognition separately. Deep learning solves problems end-to-end in a single pass, like YOLO outputting object location and label simultaneously. For complex multi-object tasks, the end-to-end deep learning approach is usually more appropriate than a decomposed ML pipeline.
How does supervised learning compare to reinforcement learning?
Supervised learning trains on a labeled dataset where each input has a known output, solving regression and classification problems. Reinforcement learning has no predefined dataset at all — an agent learns through trial and error by performing actions in an environment and observing rewards. Use supervised learning when labels exist; use reinforcement learning for sequential decision-making like self-driving cars or AlphaGo.
How does this framework compare to jumping straight into a Kaggle notebook?
A Kaggle notebook often starts mid-workflow with a chosen model, skipping problem classification and objective definition. This framework front-loads those decisions — locating the AI stage, matching the learning paradigm to your label structure, and defining the target variable precisely — so you avoid picking the wrong algorithm family or misjudging data volume before you write any code.
// Advanced
What are the ethical and regulatory concerns I should document before deploying?
Document fairness, data privacy, and regulatory compliance risks, especially if the model was trained on incomplete or biased datasets. In cybersecurity, healthcare, and financial domains, incomplete training data can cause harmful predictions. Confirm the output type matches expectations, record confidence levels, and flag any interpretability limitations of black-box models for stakeholders.
How does data volume govern whether I should use deep learning?
Data volume is the deciding factor: deep learning outperforms classical ML only when data volume is large, and it requires GPU-intensive computation. On small datasets, classical ML wins and runs on low-end machines. Assess your dataset size and hardware budget early — this single principle prevents both underperformance and wasted infrastructure spend.
When should I use the Apriori algorithm instead of K-Means?
Use the Apriori algorithm for association analysis — finding item co-occurrence patterns, most commonly in market basket analysis. Use K-Means for clustering, where you group data points into K clusters based on feature similarity. Both are unsupervised, but Apriori answers 'what goes together' while K-Means answers 'what natural segments exist.'
How do I plan infrastructure and timelines for a deep learning project?
Confirm GPU availability before building, since large neural networks require GPU-intensive computation. Expect significantly longer training time — potentially weeks from scratch. Budget infrastructure and timelines accordingly, and consider whether transfer learning or a classical ML alternative could meet your accuracy needs without the heavy compute cost.