8.3-training-neural-networks.md

8.3 Training neural networks

🌳 Tip 🌳
For more tips on training neural networks, check out:

• A Recipe for Training Neural Networks (Karpathy 2019)
• NLP's Clever Hans Moment has Arrived (Heinzerling 2019): an excellent writeup on trying to understand what exactly your neural network learns, and techniques to ensure that your model works correctly with textual data.
• An overview of gradient descent optimization algorithms (Ruder 2016)

41. [E] When building a neural network, should you overfit or underfit it first?
42. [E] Write the vanilla gradient update.
43. Neural network in simple Numpy. 26. [E] Write in plain NumPy the forward and backward pass for a two-layer feed-forward neural network with a ReLU layer in between. 27. [M] Implement vanilla dropout for the forward and backward pass in NumPy.
44. Activation functions. 28. [E] Draw the graphs for sigmoid, tanh, ReLU, and leaky ReLU. 29. [E] Pros and cons of each activation function. 30. [E] Is ReLU differentiable? What to do when it’s not differentiable? 31. [M] Derive derivatives for sigmoid function $$\sigma(x)$$ when $$x$$ is a vector.
45. [E] What’s the motivation for skip connection in neural works?
46. Vanishing and exploding gradients. 32. [E] How do we know that gradients are exploding? How do we prevent it? 33. [E] Why are RNNs especially susceptible to vanishing and exploding gradients?
47. [M] Weight normalization separates a weight vector’s norm from its gradient. How would it help with training?
48. [M] When training a large neural network, say a language model with a billion parameters, you evaluate your model on a validation set at the end of every epoch. You realize that your validation loss is often lower than your train loss. What might be happening?
49. [E] What criteria would you use for early stopping?
50. [E] Gradient descent vs SGD vs mini-batch SGD.
51. [H] It’s a common practice to train deep learning models using epochs: we sample batches from data without replacement. Why would we use epochs instead of just sampling data with replacement?
52. [M] Your model’ weights fluctuate a lot during training. How does that affect your model’s performance? What to do about it?
53. Learning rate. 34. [E] Draw a graph number of training epochs vs training error for when the learning rate is: 1. too high 2. too low 3. acceptable. 35. [E] What’s learning rate warmup? Why do we need it?
54. [E] Compare batch norm and layer norm.
55. [M] Why is squared L2 norm sometimes preferred to L2 norm for regularizing neural networks?
56. [E] Some models use weight decay: after each gradient update, the weights are multiplied by a factor slightly less than 1. What is this useful for?
57. It’s a common practice for the learning rate to be reduced throughout the training. 36. [E] What’s the motivation? 37. [M] What might be the exceptions?
58. Batch size. 38. [E] What happens to your model training when you decrease the batch size to 1? 39. [E] What happens when you use the entire training data in a batch? 40. [M] How should we adjust the learning rate as we increase or decrease the batch size?
59. [M] Why is Adagrad sometimes favored in problems with sparse gradients?
60. Adam vs. SGD. 41. [M] What can you say about the ability to converge and generalize of Adam vs. SGD? 42. [M] What else can you say about the difference between these two optimizers?
61. [M] With model parallelism, you might update your model weights using the gradients from each machine asynchronously or synchronously. What are the pros and cons of asynchronous SGD vs. synchronous SGD?
62. [M] Why shouldn’t we have two consecutive linear layers in a neural network?
63. [M] Can a neural network with only RELU (non-linearity) act as a linear classifier?
64. [M] Design the smallest neural network that can function as an XOR gate.
65. [E] Why don’t we just initialize all weights in a neural network to zero?
66. Stochasticity. 43. [M] What are some sources of randomness in a neural network? 44. [M] Sometimes stochasticity is desirable when training neural networks. Why is that?
67. Dead neuron. 45. [E] What’s a dead neuron? 46. [E] How do we detect them in our neural network? 47. [M] How to prevent them?
68. Pruning. 48. [M] Pruning is a popular technique where certain weights of a neural network are set to 0. Why is it desirable? 49. [M] How do you choose what to prune from a neural network?
69. [H] Under what conditions would it be possible to recover training data from the weight checkpoints?
70. [H] Why do we try to reduce the size of a big trained model through techniques such as knowledge distillation instead of just training a small model from the beginning?

---

This book was created by Chip Huyen with the help of wonderful friends. For feedback, errata, and suggestions, the author can be reached here. Copyright ©2021 Chip Huyen.