AT-320
Free AI+ Architect Practitioner Practice Test
12 exam-style questions with full explanations — no sign-up. Score yourself, then close your gaps with the full course.
Exam fee ~$24910 exam domainsLevel Intermediate2 timed practice tests in the course
✅ Free practice — no sign-up📝 Real exam-style questions💡 Detailed explanations💸 30-day money-back via Udemy
An architect is training a 50-layer deep neural network for image classification using sigmoid activation functions in every hidden layer. During training, the loss plateaus almost immediately and the model shows no improvement across many epochs. Inspection reveals that gradients in the early layers are near zero. Which solution BEST addresses this issue?
Sigmoid saturates and produces derivatives near zero for large positive or negative inputs; stacking many sigmoid layers compounds this into the vanishing gradient problem. ReLU has a constant gradient of 1 for positive inputs, allowing gradients to propagate through many layers without shrinking to zero.
A company is training a deep CNN for image classification. During training, the loss decreases rapidly for the first few epochs but then begins to oscillate wildly and occasionally spikes to very high values, preventing convergence. The learning rate is currently fixed at 0.1. Which approach BEST resolves this issue?
A warmup phase gradually raises the learning rate from a small value, preventing early instability, while gradual decay afterward reduces step size as the model approaches a minimum, stabilizing convergence.
A team is building a sentiment classifier for customer reviews averaging 500+ words. Using a vanilla RNN, the model performs poorly whenever the sentiment-bearing phrase appears early in the review and the final prediction depends on it. Which architecture change BEST addresses this?
Vanilla RNNs suffer from vanishing gradients, so information from early timesteps decays before reaching the output. LSTM gating mechanisms (forget, input, output gates) regulate gradient flow and allow the cell state to carry relevant information across hundreds of timesteps, directly solving this problem.
A healthcare startup has a dataset of only 500 labeled medical images and needs high classification accuracy for a diagnostic support tool. Which approach BEST meets these requirements?
Transfer learning with a pretrained ResNet-50 leverages features learned from millions of ImageNet images, requiring far less labeled data to reach high accuracy on a small 500-image dataset.
An architect is evaluating a fraud-detection model trained on transaction data where only 0.5% of transactions are fraudulent. Which metric BEST evaluates the model's performance for this architecture decision?
With severe class imbalance (0.5% positive), Precision-Recall AUC focuses evaluation on the minority (fraud) class and is not inflated by the large number of easy true negatives, making it far more informative for architecture decisions than threshold-free metrics that account for the negative class.
A fintech company must score credit card transactions for fraud as they occur, requiring sub-100ms response time and the ability to handle thousands of transactions per second. Which inference deployment pattern BEST meets these requirements?
Fraud detection with a hard sub-100ms latency requirement and high per-second throughput demands synchronous real-time serving with an autoscaling API endpoint, not periodic processing.
A fintech company deploys a credit-scoring model trained on historical loan data. Post-deployment auditing reveals the model approves loans for one demographic group at a significantly higher rate than another, even after controlling for income and credit history. Which action BEST addresses this issue in the AI architecture?
Fairness-aware re-weighting or adversarial debiasing directly targets the discriminatory pattern in training, and validating with a disparate impact ratio provides measurable evidence the bias has been reduced.
An enterprise needs an LLM-based assistant to answer employee questions using internal policy documents that are updated daily. The architecture must minimize hallucinations and avoid the cost of retraining the model whenever documents change. Which solution BEST meets these requirements?
RAG decouples knowledge updates from model training: new or changed documents are re-embedded and indexed in the vector store, and the LLM retrieves current, relevant context at inference time, reducing hallucination without retraining.
A research team proposes adopting a newly published transformer variant that reports a 2% accuracy improvement on a public benchmark. Before integrating it into the production architecture, which action BEST reflects evidence-based architecture practice?
Evidence-based architecture practice requires independent reproduction on representative data and statistical validation before any production change, since benchmark gains may not generalize or be significant.
An architect is finalizing a capstone design for a multimodal customer-support system that must transcribe speech, classify intent (NLP), analyze uploaded product images (CV), and generate a response using an LLM, while meeting a regulatory requirement for explainability and bias monitoring. Which architecture BEST satisfies all requirements?
Modular pipelines with a governed orchestration layer allow per-stage logging, bias metrics, and explainability, which is the pattern taught across the infrastructure, evaluation, and responsible AI domains for complex multimodal systems.
A team is designing the output layer of a neural network for a binary classification task (spam vs. not spam). Which combination of output activation function and loss function is BEST suited for this task?
Sigmoid squashes the single output neuron into a (0,1) probability range, and binary cross-entropy is the loss function mathematically derived to optimize that probability against a binary (0/1) label.
A deep learning engineer trains a model that achieves 99% training accuracy but only 70% validation accuracy, indicating severe overfitting. Which single approach BEST addresses this problem?
Combining dropout, L2 weight decay, and data augmentation directly attacks overfitting by constraining model complexity, penalizing large weights, and increasing effective training data diversity.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →AT-320 exam — quick answers
How much does the AT-320 exam cost?
The exam fee is approximately $249 and varies by region — confirm current pricing with the certification vendor before you book.
What topics are on the exam?
It covers 10 domains: Fundamentals of Neural Networks (10%), Neural Network Optimization (10%), Neural Network Architectures for NLP (10%), Neural Network Architectures for Computer Vision (10%), Model Evaluation and Performance Metrics (10%), AI Infrastructure and Deployment (10%), AI Ethics and Responsible AI Design (10%), Generative AI Models (10%), Research-Based AI Design (10%), Capstone Project and Course Review (10%). The full course has a dedicated chapter, lab and practice-test coverage for each.
Is this practice test really free?
Yes — all questions on this page are free with explanations and no sign-up. The paid Udemy course adds two full-length timed exams, video lessons and hands-on labs.
Will this prepare me for the real exam?
The questions mirror the real exam's style and are mapped to the official domains. This is exam-focused preparation — combine the free test with the full course's timed simulations to gauge your readiness.