Machine Learning Engineer Interview Questions

Satyam C. - "TF-IDF CONCEPT EXPLANATION AND INTUITION BUILDING: TF-IDF is a measure that reflects the importance of a word in the document relative to a collection of documents. Its full form is Term Frequency - Inverse Document Frequency. The term TF indicates how often a term occurs in a particular document. It is the ratio of count of a particular term in a document to the number of terms in that particular document. So, the intuition is that if a term occurs frequently in a single documen"See full answer

Anmol R. - "The example given is wrong. The 2nd test case should have answer 3, as we can get to 6 by using 3 coins of denomination 2."See full answer

Murali M. - "The rule doesn't work the other way around. If the array is smaller than n, it can still have duplicates. Eg: n=10 , arr = [3,3]"See full answer

Xuelong A. - "I've participated in several competitions in Kaggle concerning medical images. My most recent competition deals with images of skin lesions and classifying them as either melanoma or not. I focused on fine-tuning pretrained models and ensembling them. I also like to keep track of the latest trends of computer vision research, with a focus on making models memory-efficient through model compression and interpretability."See full answer

Tiago R. - "function isPalindrome(s, start, end) { while (s[start] === s[end] && end >= start) { start++; end--; } return end <= start; } function longestPalindromicSubstring(s) { let longestPalindrome = ''; for (let i=0; i < s.length; i++) { let j = s.length-1; while (s[i] !== s[j] && i <= j) { j--; } if (s[i] === s[j]) { if (isPalindrome(s, i, j)) { const validPalindrome = s.substring(i, j+1"See full answer

Lash - "A perceptron is the most basic building block of a neural network and represents a single-layer binary classifier."See full answer

Madhur K. - "I have worked with ML when I was a PM for Customer Due Diligence and it required recalibration to improve the matching logic (for screening and Proof of business matching with Google) and threshold analysis. In this scenario we had to extract past True positive, False positive decisions from operations and run threshold analysis to come up with a new threshold to match better and remove the noise i.e. False positives. Taking this new threshold for matching, the ML algorithm is receiving feedback"See full answer

Matt M. - "Given the dataset does not contain many labels, it implies we cannot directly use supervised learning. I would ask more about the type of dataset we are given. Is it images, text, etc? This may inform the types of transformations we do the dataset. I can see two approaches to training Given the labels we do have, we can find a method to generate labels for the other unlabeled data. This likely will introduce some error since they may not be true labels, but it at least allows processing the"See full answer

Shibin P. - "Effective loss functions for computer vision models vary depending on the specific task, some commonly used loss functions for different tasks: Classification Cross-Entropy Loss:Used for multi-class classification tasks. Measures the difference between the predicted probability distribution and the true distribution. Binary Cross-Entropy Loss:Used for binary classification tasks. Evaluates the performance of a model by comparing predicted probabilities to the true binary labe"See full answer