Questions
Questions

11785/11685/11485 Quiz-14

Single choice

An exhausted TA for CMU’s Introduction to Deep Learning (11-785) has reached their limit. After weeks of spelunking through Piazza threads filled with cryptic stack traces, malformed attention plots, and variable names like dWhx_v3_final_fr, they’ve decided enough is enough. Rather than manually traversing the combinatorial space of student mistakes in HWP2, the TA sets out to automate out some of the pain. To do so, they scrape every relevant Piazza post from the past N semesters (where N ≈ too many), including: Descriptions of bugs (both coherent and not) Fragments of semi-functional student code TA replies ranging from thoughtful diagnostics to "try turning it off and on again" Post metadata (HW category, resolution status etc..) Now armed with this glorious mess of historical bug data, the TA wants to train a deep learning system capable of: Bug Detection: Detecting whether a post involves a bug Bug Classification: Predicting the likely type of bug (e.g., “you forgot to mask” or “why is your loss negative?”) Bug Retrieval: Retrieving similar past bug reports The dream? Never having to answer “Why is my CER 600?” at 2am again. For the following question, assume the TA is able to construct some representative dataset with the following properties: Each Piazza post is labeled as either bug or not bug. For posts labeled as bug, one mutually exclusive bug type is provided. For each bug-labeled post, one or more aligned fix explanations are available.   Which of the following modeling approaches is best suited to detect whether a Piazza post describes a bug or not? (Select all that apply) 

Options
A.An Encoder-Decoder Transformer, trained on aligned (post → fix) examples where posts that don’t trigger generation are assumed to be bug-free.
B.A pre-trained bidirectional Transformer encoder (e.g., BERT) trained on a large corpus of unlabeled text via masked language modeling, with a classification head fine-tuned on the using binary cross-entropy.
C.An RNN encoder with max-pooling across time and softmax classifier trained over C known bug categories using cross-entropy loss.
D.A pairwise CNN encoder architecture trained with contrastive loss to minimize distance between similar bug embeddings.
E.A reconstruction autoencoder trained on all posts labeled as a bug, where thresholding on reconstruction loss is used as a proxy for bug detection.
F.A clustering algorithm applied to the latent space of a Variational Autoencoder (VAE) trained on bug-labeled posts, where each resulting cluster is interpreted as a latent bug category.
G.An MLP based single softmax classifier over C+1 classes where the C bug classes are augmented with a “no bug” class.
View Explanation

View Explanation

Verified Answer
Please login to view
Step-by-Step Analysis
The question asks which modeling approaches are best suited to detect whether a Piazza post describes a bug, with the additional setup that each post is labeled bug or not bug, and for bug posts a specific bug type is provided. Option 1: An Encoder-Decoder Transformer, trained on aligned (post → fix) examples where posts that don’t trigger generation are assumed to be bug-free. This approach resembles a generative retrieval/generation setup rather than a discriminative binary detector. It presumes you’ll generate a fix for every post, and uses the generation signal to infer bugs, which is indirect and can be brittle for detection. It also relies on a problematic assumption that non-generation implies bug-free, which can lead to false negatives/positives if the model is uncertain or if non-bug posts resemble fixable content. Option 2: A pr......Login to view full explanation

Log in for full answers

We've collected over 50,000 authentic exam questions and detailed explanations from around the globe. Log in now and get instant access to the answers!

Similar Questions

A company wants to build a system that automatically classifies incoming customer emails as either "complaint," "inquiry," or "feedback" based on the words in the message. They are looking for a simple, fast solution that works well with large amounts of text data and doesn't require extensive training time. Which machine learning approach is most appropriate for this task?

The supervised machine learning approach to sentiment analysis is based on:

Text classification utilizes a unique set of classification algorithms, distinct from algorithms used in predictive analytics with numeric data:

Which of the following is a better process for building a text classification model?

More Practical Tools for Students Powered by AI Study Helper

Join us and instantly unlock extensive past papers & exclusive solutions to get a head start on your studies!