Machine Learning Interview Questions

Varun G. - "Constrain the requirements (The final product should be somewhat like Neptune AI and also enable for A/B testing) Do some back of the envelope calculations Start explaning your system with appropriate architecture diagrams Optimize for scale Answer any questions Wrap-up"See full answer

Alok S. - "Functional requirements: user can send an input and wait for the result Group up to 100 individual requests in to single GPU The system should should send results back to the user who requested it when done Non functional requirements: Minimize the waiting between two batches of execution/ reduce idle time error message if a batch faiils Scale to support multiple GPUs Core Entities: Request Batch Result API Design: POST /predict -> {requestid: "", response: ""} req"See full answer

Ajay P. - "Designing an evaluation framework for ads ranking is crucial for optimizing the effectiveness and relevance of ads displayed to users. Here's a comprehensive framework that you can use: Define Objectives and Key Performance Indicators (KPIs):** \\Click-Through Rate (CTR):\\ The ratio of clicks to impressions, indicating the effectiveness of an ad in attracting user attention. \\Conversion Rate:\\ The ratio of conversions (e.g., sign-ups, purchases) to clicks, measuring how well"See full answer

Alex N. - "Since question asks about pipeline. I assume the question is about metrics across many dimensions not just prediction Model performance. For the ML Model: I can use accuracy, precision, recall, F1 if it is classification model. In case it is regression model RMSE is good metric for many problems. Data: ML system needs good quality data. The system has to track missing data rate. Distribution of features, if there is no drift from original feature distributions during the training. Pipeline h"See full answer

Anonymous Partridge - "We will have to use a second more powerful LLM Model to validate the answers. LLM as a judge"See full answer

Alex N. - "The adjusting context window size in LLM change trade off between reasoning capability, accuracy, computation cost. It influence how attention mechanist allocate resources across the input. Longer context window let it you input greater number of words and have more context to generate proper next token. However llms have lost in the middle issue. They remember the beginning of text and end of text but lost information located in the middle of long input. Another problem is Attention Dilution."See full answer

Anchal V. - "supervised learning: model is trained on the labeled data unsupervised learning: no labels provided - model learns by finding patterns , structure and groupings in the data. Semi-supervised learning: use small set of labels to guide learning for the larger pool of unlabeled data. reinforcement learning: leans by interacting with students the environment, receives reward and penalties based on actions self supervised: no labelled data . The model makes its own practice problems by"See full answer

Kartikeya N. - "C : Okay. So I would want to start with knowing what is the product for which we have to build a recommendation system. I : This is a photo sharing product. C : Okay. So is this something on the lines of Instagram? I : Yes C : Okay. And are we a new product co or we have some current product built already? I : You can assume yourself. C : Okay. Is there any demography or country we are targeting? I : No, this is a global product C : Okay. So, the biggest goal of any product recommendation system"See full answer

Hardik saurabh G. - "Hallucinations are evaluated by measuring how often generated outputs contain information that is not supported by trusted sources. what hallucination means in context: Intrinsic hallucination: contradicts provided context Extrinsic hallucination: introduces unsupported facts Fabrication: confidently incorrect answers"See full answer

Kapil D. - "Overall a very good summary of ranking problems. I wish there are more details in the post on depth of modeling. One can talk more about advance ML architectures like DCN, BertRec etc"See full answer

Sai vuppalapati M. - "At a high level, the core challenge here revolves around building an effective recommendation algorithm for news. News is an inherently diverse category, spanning various topics and catering to a wide array of user types and personas, such as adults, business professionals, general readers, or specific cohorts with unique interests. Consequently, developing a single, one-size-fits-all recommendation algorithm is not feasible. To enhance the personalization of the news recommendation algorithm,"See full answer

Dinar M. - " Thanks a lot for showing us how a recommender system can be build. I see it was proposed to use Collaborative filtering which is user - item matrix having dimension N * M (where N - number os users and M - number of songs). Though, it was explained how it gonna be built, it is still unclear how all users and songs features are going to be used. In that matrix we have values in cell (lets say i, j) like 1 - a specific user (i) clicked on song (j) when it was recommended or it is 0 when the user"See full answer

Megha V. - "Gradient Descent is an optimisation strategy used in several supervised learning models. It is the technique for finding the optimum solution of an objective function. Typically, for a linear regression use case, it is used to find the weights and bias that produce the lowest loss. It involves computing the partial derivative of the objective function with respect to the weight and bias vectors. To find the optima of the function, the derivative is equated to 0, and iteratively the weight and b"See full answer

Clayton P. - "The interviewer hinted that a two-tower recommender system might be a suitable approach, using user history to embed users and pages separately and train on view or interaction data. Instead, I proposed a different approach that I felt was more aligned with how knowledge is structured in Confluence: I designed a system using a graph database to model the relationships between Confluence pages. Each page is a node, and edges represent content-based references. For example, when one article"See full answer