What to Say When You Don't Know an ML Interview Answer

In an ML fresher interview in India, blanking on a question damages you less than responding with a flat “I don’t know” and offering nothing else.

The distinction matters because ML interviewers are not just testing recall. They’re evaluating how you reason when the answer isn’t obvious. A flat non-response tells them nothing about your thinking. A structured partial answer, even an imperfect one, tells them quite a lot.

There are three response patterns that work. Each fits a different situation. Each gives the interviewer something to probe.

Why the Flat “I Don’t Know” Fails

It isn’t the honesty that hurts. Honest candidates do fine in ML interviews. What fails is the conversational dead end.

When you say “I don’t know” and stop, the interviewer has two uncomfortable options: move on and note that you had nothing, or try to rescue the conversation with leading hints. Neither is good for you.

Compare that to this:

“I haven’t memorised this formula exactly, but I know that L2 regularisation adds a penalty proportional to the squared weights to the loss. My understanding is that it shrinks weights toward zero without forcing most of them to exactly zero, which is what L1 does. I’d need to work through the gradient derivation to be precise.”

That response shows conceptual understanding, correct framing, and honest calibration of confidence. The interviewer can probe any of those three layers. You’ve given them five ways to continue the conversation instead of zero.

Response Pattern 1: The Partial-Knowledge Script

Use this when you understand what the concept does but can’t recall the exact formula or implementation detail.

The structure is:

• What the concept is solving: the problem it addresses
• What you know with confidence: the conceptual layer you can stand behind
• Where your confidence drops: name the specific gap
• Your best reasoning from there: a conclusion even without the derivation

Example: L2 regularisation math

• “L2 regularisation is addressing overfitting by penalising large weights.”
• “The penalty is proportional to the squared magnitude of the weights. Adding it to the loss pushes the optimiser to prefer smaller weights during training.”
• “I’d need to work through the gradient derivation step by step to give you the exact update rule.”
• “But I know the practical effect is that weights shrink but rarely reach exactly zero, which is why L2 tends to keep all features in the model while L1 can zero some out.”

That is a complete, honest answer. It shows you understand the mechanism even without the formula memorised.

The overfitting and regularisation article has the full derivation if you want to close that specific gap before your interview round.

Response Pattern 2: The First-Principles Script

Use this for questions you’ve never encountered directly but where adjacent concepts give you a foothold.

The structure is:

• Say you’re going to reason from first principles
• Name the adjacent concepts you’re using as starting points
• Build the argument step by step, out loud

Example: Comparing RMSprop and Adam optimisers

Say you’ve never directly compared these two, but you know both are adaptive learning rate methods:

• “I haven’t compared these directly, so let me reason from what I know.”
• “Both adjust the learning rate per parameter based on recent gradient history. RMSprop keeps a running average of squared gradients and divides the learning rate by the root of that average.”
• “Adam adds momentum on top — it tracks a moving average of gradients themselves, not just their squares. That extra term helps Adam keep moving through noisy or sparse gradient regions.”
• “So my reasoning is: for a well-behaved, dense dataset, RMSprop and Adam probably converge similarly. For noisier problems or sparse inputs, Adam’s momentum term would help more. I’d start with Adam in practice.”

That last bullet is a practical engineering conclusion, which is exactly what interviewers in ML system design rounds are looking for. The Google Machine Learning Crash Course covers optimiser families if you want the formal definitions alongside this reasoning pattern.

Response Pattern 3: The Honest Redirect

Use this when you genuinely haven’t covered the topic and can’t build from adjacent concepts.

The structure is:

• Acknowledge the gap directly, without excessive hedging
• Pivot to adjacent knowledge you do have
• Ask a question that re-opens the conversation

Example: Attention score computation in transformers

If you haven’t studied the attention mechanism in detail:

• “I haven’t gone deep into the transformer attention mechanism yet.”
• “What I have covered is the general encoder-decoder architecture and the problem attention was designed to solve: giving the decoder dynamic access to all encoder states, not just the final hidden state.”
• “Could you tell me which part you’re probing? I’d like to reason through it even if I don’t have the formula exact.”

That’s honest and opens a dialogue. Most interviewers will give a hint or redirect. It also sets up a strong follow-up: if the interviewer explains the attention score formula, you can connect it back to what you said about encoder-decoder access, showing you’re absorbing and integrating in real time.

If attention questions come up often in your target rounds, the transformer explainer covers the attention computation in the structured format interviewers expect.

Your own project is also legitimate redirect territory. The project walkthrough article covers exactly this: when you hit a wall on a theoretical question, redirecting to a concrete decision you made in a real project is a credible pivot, not a deflection.

The Five ML Concepts Where Freshers Most Often Blank

These topics appear most often in Indian AI/ML fresher interview rounds and produce the highest blank-out rate.

Concept	What usually trips freshers	Partial-knowledge anchor
Bias-variance tradeoff	Definition is memorised, worked example breaks down	”High bias means the model’s assumptions don’t fit the data — underfitting. High variance means it’s too sensitive to training noise — overfitting.”
Overfitting and regularisation	Regularisation math	”The penalty term discourages large weights. L1 pushes weights toward zero, L2 pushes weights toward small but non-zero.”
Evaluation metrics beyond accuracy	When to use F1 vs AUC-ROC	”For imbalanced classes, accuracy misleads. Precision and recall trade off against each other — F1 balances them. AUC-ROC tells you how well the model separates classes across all thresholds.”
Attention in transformers	Exact Q, K, V computation	”Attention gives the decoder dynamic, weighted access to all encoder states. The weights are learned, not fixed.”
Gradient descent variants	Adam vs SGD vs RMSprop details	”Adam adds momentum to adaptive learning rates. SGD is simpler but needs careful learning rate tuning. Adam is the practical default for most problems.”

For evaluation metrics, the scikit-learn model evaluation documentation is the clearest reference: it maps which metric fits which problem type with concrete examples, including the precision-recall tradeoff for imbalanced datasets.

The bias-variance tradeoff article has ready-to-use 60-second and 2-minute scripted answers for one of the most common blank topics.

What These Three Patterns Have in Common

All three share one mechanism: narration. You say what you know, say where your confidence stops, and say where your reasoning is headed. Interviewers in ML rounds can distinguish between a candidate who knows the material but can’t recall a formula and a candidate who doesn’t know the material. The gap between them, more often than not, is whether you narrate.

Once you’ve internalised these patterns, the next step is making sure your coverage of the core ML concepts is broad enough that the partial-knowledge script applies to most questions you face. The 2026 AI roadmap for Indian engineering students maps those concepts by priority and fits them into a timeline that works within a 6-month placement window.