Deep Learning flashcards that match how you actually study
Whether you are prepping for exams or building long-term knowledge, Deep Learning rewards retrieval practice—not rereading. NoteFren converts your handwritten notes, slides, and PDF text into clean Q&A flashcards so you can review Deep Learning with spaced repetition in minutes, not hours.
Studying Deep Learning with flashcards
Deep learning focuses on neural networks: layers and activations, backpropagation, loss functions, optimizers, and specialized architectures like CNNs, RNNs/LSTMs, and Transformers. Students must hold in memory a large set of interlocking details — the shapes of tensors through a layer, what softmax versus sigmoid outputs, how gradients flow, and why techniques like batch normalization, dropout, or residual connections exist. The math (chain rule, matrix dimensions) and the vocabulary of hyperparameters are where recall commonly breaks down.
Active recall suits this because most exam and interview questions are pointed: "What problem does dropout solve?" or "Why do vanishing gradients happen in deep sigmoids?" Spaced repetition keeps architecture components and their motivations distinct instead of merging into vague intuition. Write cards that pair a technique with the specific problem it addresses, cards that ask for output shapes given an input and layer parameters, and cards contrasting close relatives (Adam vs SGD with momentum, LSTM vs GRU). Photographing a hand-drawn network diagram into NoteFren and generating cards helps you drill the forward and backward pass without redrawing it each time.
Key topics to turn into flashcards
Activation functions
Card ReLU, sigmoid, tanh, and softmax with their ranges, derivatives, and the problems they cause or solve (dead ReLUs, vanishing gradients).
Backpropagation and gradients
Test the chain rule through a layer, what vanishing and exploding gradients are, and how they are mitigated.
Loss functions
Front names cross-entropy, MSE, or hinge loss; back gives the formula and when to use it (classification vs regression).
Convolutional networks
Card kernel size, stride, padding, and the formula for output spatial dimensions, plus what pooling does.
Regularization techniques
Pair dropout, weight decay, batch norm, and early stopping each with the specific overfitting or training problem it targets.
Transformers and attention
Cover query-key-value attention, why positional encodings are needed, and what multi-head attention adds.
Study tips
- Tip 1
Chunk by topic
Split Deep Learning into small decks—one per lecture, chapter, or concept—so reviews stay fast and focused.
- Tip 2
Answer before you flip
Say the answer out loud or jot a keyword before revealing the card. Active recall beats passive recognition every time.
- Tip 3
Schedule reviews
Let spaced repetition surface Deep Learning cards right before you would forget them. Cramming alone rarely sticks.
- Tip 4
Use mistakes as data
Tag or star misses and revisit them first next session—your weak spots are where the most points hide.
Common mistakes to avoid
Treating architectures as black boxes
Naming a layer is not knowing what it computes. Card the exact operation and tensor shapes so you can trace a forward pass.
Memorizing formulas without the motivation
Reciting the Adam update is useless if you cannot say why it beats plain SGD. Always card the problem each method solves.
Ignoring dimension bookkeeping
Shape mismatches sink real implementations. Practice computing output dimensions for conv and dense layers on cards.
Sample Deep Learning flashcards
A few cards students actually use for Deep Learning. Save a deck in NoteFren to study them with spaced repetition.
Tap a card to check yourself.
1
What does ReLU compute?
Show answerHide answer
max(0, x): positive inputs pass through and negative ones become zero.
2
Why can sigmoid activations slow training in deep networks?
Show answerHide answer
They saturate, so gradients shrink toward zero as they pass back through many layers.
3
Which loss suits multi-class classification?
Show answerHide answer
Cross-entropy loss on softmax outputs.
4
What does dropout do during training?
Show answerHide answer
Randomly zeroes a fraction of activations so the network cannot rely on any single unit.
5
What is the output size of a 32×32 input with a 5×5 kernel, stride 1 and no padding?
Show answerHide answer
28×28, since 32 − 5 + 1 = 28.
6
What does batch normalization do?
Show answerHide answer
Normalizes a layer's inputs over each mini-batch, then rescales and shifts them with learned parameters.
7
What is scaled dot-product attention?
Show answerHide answer
softmax(QKᵀ/√d_k)V: values weighted by how well each key matches the query.
8
Why do transformers need positional encodings?
Show answerHide answer
Self-attention ignores token order, so position information has to be added.
Frequently asked questions
Yes. NoteFren turns your notes and photos into smart flashcards with spaced repetition and active recall—ideal for mastering Deep Learning without retyping everything.
NoteFren is an iOS app built for focused study sessions. Check the App Store listing for the latest connectivity and sync details.
Absolutely. Every card can be edited, merged, or deleted so your deck matches exactly what you need to learn.
Related subjects & guides
Create your own flashcards
Upload your notes or a PDF and get a deck you can study — free, in your browser.
Works in your browser — no download needed.