Middle AI/ML Engineer — AI/ML engineer interview questions

What causes vanishing and exploding gradients, and how do you deal with them?

In a deep network, gradients are multiplied together across many layers during backpropagation, so if those factors are consistently below 1 the gradient shrinks toward zero by the time it reaches early layers, and if they are above 1 it blows up. ReLU-family activations, careful weight initialization (Xavier, He), batch or layer normalization, and residual connections all address this — which is part of why ResNets and transformers can be trained so much deeper than plain feedforward nets.

What is transfer learning, and when does it help?

Transfer learning takes a model pretrained on a large, general dataset and adapts it to a smaller, more specific task, either by fine-tuning some layers or by using it as a frozen feature extractor. It helps most when your own dataset is too small to train a good model from scratch — which is most real projects — because the pretrained model already learned generic features like edges or syntax that transfer across tasks.

How do CNNs, RNNs and transformers differ in what they are good at?

CNNs use convolutional filters that exploit spatial locality, which is why they dominate image tasks. RNNs process sequences step by step and carry a hidden state forward, which made them the default for text and time series until their sequential nature made them slow to train and bad at long-range dependencies. Transformers process the whole sequence in parallel using attention, which is both faster to train and better at long-range dependencies, and is why they now dominate NLP and are increasingly used for vision too.

Walk through a typical PyTorch or TensorFlow training loop.

Load a batch, run the forward pass to get predictions, compute the loss against the ground truth, call backward to compute gradients via autograd, then step the optimizer to update the weights and zero the gradients before the next batch. Around that core loop sits validation on a held-out set every epoch, checkpointing the best model, and usually a learning rate scheduler that decays the rate as training progresses.

How do you approach hyperparameter tuning?

Grid search is exhaustive but expensive and wasteful when some parameters barely matter. Random search often finds a good configuration faster because it does not waste trials on unimportant dimensions. Tools like Optuna use Bayesian optimization to actively pick the next configuration based on past results, which converges faster than either — but the real skill is narrowing the search space first using domain knowledge instead of tuning blindly.