第 23 / 41 页
非代码面试题
显示 20 / 811 道匹配题目
答题状态：未尝试未正确已正确
ID题目领域难度题型进度权限
4336Why CNN Can WinWhy can a modest CNN beat a larger Transformer on a small-data task whose label depends mainly on short local patterns?机器学习中等essay未尝试面试订阅4337Why RNN Still MattersWhy might an RNN still be the practical choice for a production event-stream model even if Transformers benchmark better offline?机器学习中等essay未尝试面试订阅4338When Attention Earns Its CostWhat kind of task structure makes the quadratic cost of attention worth paying?机器学习中等essay未尝试面试订阅4339Why Architecture Mismatch HurtsWhy can architecture mismatch dominate parameter count when performance is poor?机器学习中等essay未尝试面试订阅4340Hybrid Versus PureWhen is it more sensible to consider a hybrid architecture instead of insisting on a pure CNN, pure RNN, or pure Transformer?机器学习中等essay未尝试面试订阅4356Trade Ranking Versus Probability QualityModel A has slightly higher ROC AUC, but its probabilities are visibly overconfident; Model B has slightly lower ROC AUC but much better calibration. If you need a probability to size positions according to expected payoff, which model is usually safer?机器学习中等essay未尝试面试订阅4357Severe Imbalance Metric ChoiceYou are screening rare fraud cases with a base rate below 1%. Why is PR analysis usually more informative than ROC analysis for the operational discussion?机器学习中等essay未尝试面试订阅4358Same AUC Different ReliabilityTwo models have essentially the same ROC AUC. One is well calibrated, the other places most probabilities near 0 or 1. What should be your first concern before picking the latter?机器学习中等essay未尝试面试订阅4359Thresholds Need ContextWhy is it a mistake to ask for the 'best threshold' without specifying a cost ratio or downstream resource constraint?机器学习中等essay未尝试面试订阅4360Calibration After ShiftA model was calibrated on last year's data, but the event base rate has clearly shifted this year. What is the first calibration question you should revisit?机器学习中等essay未尝试面试订阅4361Before Comparing CurvesBefore you compare ROC and PR curves across models, what dataset property should you check first?机器学习中等essay未尝试面试订阅4362Before RecalibratingA model looks poorly calibrated. What should you check first before concluding the model itself is broken?机器学习中等essay未尝试面试订阅4363Before Picking a ThresholdWhat is the first quantity you should pin down before optimizing a classification threshold?机器学习中等essay未尝试面试订阅4364Before Trusting AUCAUC improved a little after retraining. What should you ask first before declaring the new model practically better?机器学习中等essay未尝试面试订阅4365Calibration Is Not EnoughWhy should you hesitate before selecting a perfectly calibrated model that has weak ranking power?机器学习中等essay未尝试面试订阅4376Validation Set WorshipOnly 1% of cases are positive, and the desk can manually investigate just the top 100 alerts each day. When tuning thresholds, would PR-oriented metrics or ROC-oriented metrics deserve more emphasis first?机器学习中等essay未尝试面试订阅4377Grid Or Random SearchA feature standardization step is fitted on the full dataset before cross-validation instead of inside each fold. What is the main tuning problem with that workflow?机器学习中等essay未尝试面试订阅4378Early Stopping LeakageYou can afford only 30 evaluations, and experience suggests that only a few hyperparameters matter strongly while the rest are weak. Would grid search or random search usually deserve the first try?机器学习中等essay未尝试面试订阅4379Budget-Limited TuningData are scarce but the search space is broad, and you want an almost unbiased performance estimate after tuning. Is nested CV conceptually appropriate here despite its cost?机器学习中等essay未尝试面试订阅4380Fair Model ComparisonA researcher keeps rerunning the tuning loop until one hyperparameter setting looks best on CV by a tiny margin. What is the core risk in that behavior?机器学习中等essay未尝试面试订阅