Scribd
Upload
94 views2 pages

AI Engineer Cheat Sheet Micro1

The document is a cheat sheet for AI Engineer interviews covering key topics in deep learning, neural networks, and frameworks like PyTorch and TensorFlow. It discusses various concepts such as normalization techniques, model training, NLP methods, evaluation metrics, and deployment strategies. Additionally, it emphasizes the importance of communication, project examples, and ethical considerations in machine learning.

Uploaded by

mridsharma01
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
94 views2 pages

AI Engineer Cheat Sheet Micro1

The document is a cheat sheet for AI Engineer interviews covering key topics in deep learning, neural networks, and frameworks like PyTorch and TensorFlow. It discusses various concepts such as normalization techniques, model training, NLP methods, evaluation metrics, and deployment strategies. Additionally, it emphasizes the importance of communication, project examples, and ethical considerations in machine learning.

Uploaded by

mridsharma01
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 2

AI Engineer Interview Cheat Sheet –

Micro1
1. Deep Learning & Neural Networks
BatchNorm vs LayerNorm: BatchNorm normalizes across batch dimension; LayerNorm
normalizes across features. LayerNorm is better for RNNs/Transformers.

Backpropagation: Computes gradients using the chain rule from output to input, enabling
weight updates via gradient descent.

Vanishing/Exploding Gradients: Vanishing = small gradients; Exploding = large gradients.


Use ReLU, gradient clipping, proper initialization, LSTM/ResNet.

Dropout: Randomly disables neurons during training to reduce overfitting and improve
generalization.

CNNs vs Dense Networks: CNNs use spatial structure and weight sharing, making them
efficient for image data.

2. PyTorch / TensorFlow
Custom Model in PyTorch: Subclass nn.Module, define layers in __init__, logic in forward().
Train using loss.backward() and optimizer.step().

nn.Module vs nn.functional: Module manages layers; functional offers stateless ops for
flexibility.

PyTorch to TensorFlow: Export with torch.onnx.export(), import to TF using ONNX-TF.

TF Optimizations: Use tf.function, mixed precision, tf.data API, quantization.

Eager vs Graph Mode: Eager = immediate execution (easy debugging); Graph = optimized
static graph (better performance).

3. NLP
RNN vs LSTM vs GRU: LSTM & GRU solve RNN’s long-term dependency issues via gating;
GRU is simpler than LSTM.

Transformer: Uses self-attention and positional encoding to model sequences; highly


parallelizable.
Attention Mechanism: Focuses on relevant input parts; improves performance in
sequence tasks.

Fine-tune BERT: Add classification head, tokenize inputs, train using cross-entropy loss.

OOV Handling: Use subword tokenizers like BPE/WordPiece to reduce OOV impact.

4. Practical ML & Deployment


NLP Evaluation: Use accuracy, F1, precision, recall; BLEU/ROUGE for generation.

Text Classification Metrics: Accuracy, precision, recall, F1-score, confusion matrix.

Model Deployment: Save model, wrap in FastAPI, deploy with Docker or cloud service.

Imbalanced Data: Use class weights, resampling, F1-score; avoid accuracy bias.

Text Preprocessing: Clean, tokenize, normalize, remove stopwords, use embeddings or


transformer tokenizers.

5. Communication & Projects


Project Example: Built NLP pipeline with BERT for entity extraction in legal docs;
improved automation by 40%.

Model Selection Disputes: Use experiments, explain trade-offs, align with business goals.

Staying Updated: Follow ArXiv, attend webinars, contribute on GitHub/Kaggle.

Debugging Models: Check data, loss curve, overfitting, tune hyperparameters, try simpler
models.

Interpretability & Ethics: Use SHAP/LIME, test for bias, ensure data transparency.

You might also like