How to use artificial intelligence flashcards

A useful artificial intelligence flashcard deck should test one decision, definition, formula or mechanism at a time. It should not ask you to reproduce an entire lecture from one oversized card. For each card, answer before looking at the reverse side, then grade the answer honestly:

  • Again — you could not give the answer or made a serious error.
  • Hard — you knew the idea but missed an important condition, formula or distinction.
  • Good — your answer was accurate without prompting.
  • Easy — you answered accurately and quickly.

The deck below is designed for a computer science revision session. It covers the foundations that connect many artificial intelligence topics: learning paradigms, data splitting, evaluation, optimisation, neural networks, language models and responsible use. Say each answer aloud or write it in a sentence before revealing the answer. The aim is retrieval, not recognition.

MySummaries turns a set of artificial intelligence notes into a deck in this format, so you can work through the cards and send difficult ones back into later sessions. Start with the question shown below, reveal the answer, choose a grade, and continue through the queue.

A single study session on this topic can look like this:

Cards — Artificial Intelligence — Foundations18 due

What is the defining feature of supervised learning?

The model learns from labelled examples, using an input and its known target output for each training example.

All 18 cards
What is the defining feature of supervised learning?The model learns from labelled examples, using an input and its known target output for each training example.
What is the purpose of a validation set?To choose models or hyperparameters during development without using the held-out test set.
What is data leakage?Information unavailable at prediction time, or information from the test set, enters training and makes evaluation look better than real performance.
Define overfitting.Learning the training data too closely, including noise, so training performance is high but performance on unseen data is poor.
What does precision measure?Of the examples predicted positive, the proportion that are actually positive: TP/(TP+FP).
What does recall measure?Of the actually positive examples, the proportion the model identifies: TP/(TP+FN).
When is F1 score useful?When a single measure should balance precision and recall, especially when accuracy alone is misleading.
What is gradient descent changing during training?The model parameters, in the direction that reduces the loss, using the gradient of that loss.
What is the learning rate?The step size used to update parameters during optimisation.
What is the purpose of an activation function in a neural network?It introduces non-linearity, allowing stacked layers to represent functions more complex than a single linear transformation.
What does a convolutional layer exploit in image data?Local spatial structure and shared weights, so the same learned filter can detect a feature in different image locations.
What is an embedding?A learned numerical vector representation in which useful relationships between items can be represented by their positions or directions.
What is self-attention calculating?How strongly each token should use information from other tokens when constructing its contextual representation.
Why are transformers useful for language tasks?They use attention to model relationships between tokens and can process sequence positions in parallel during training.
What is a language model trained to predict?Usually the next token, or a related token sequence objective, from the preceding context.
What is a hallucination in a generative AI system?A fluent output that presents unsupported or false content as if it were reliable.
What does regularisation aim to do?Reduce overfitting by discouraging unnecessarily complex solutions or parameter values.
What is the difference between a parameter and a hyperparameter?A parameter is learned from data; a hyperparameter is chosen outside the learning process, such as learning rate or tree depth.
An artificial intelligence flashcard deck with the first card open for grading and the remaining cards waiting in the queue.

The right way to study this deck

Do not treat the answers as notes to reread. For each card, attempt a complete response first. A correct answer to the precision card needs both the meaning and the formula. A correct answer to the validation-set card needs the distinction from the test set. If you only remember a related term, mark the card Again or Hard rather than rewarding recognition.

After the first pass, group the cards by the error you made:

  1. Definition error: you confused two terms, such as validation and test data.
  2. Formula error: you knew what a metric meant but could not reconstruct its denominator.
  3. Mechanism error: you knew that attention was important but could not explain what it calculates.
  4. Application error: you could define overfitting but could not identify leakage in a deployment example.

Repeat the cards you marked Again after a short break. Leave Good cards for a later interval, and do not spend the session rereading the cards you marked Easy. Spaced repetition works best when the review interval reflects the strength of the memory rather than treating every card identically.

The core behind the deck

The 18 cards are cut from a smaller set of ideas that should remain available even when you forget details. This core is useful before an exam, interview or practical model-building task. Try to recite each item without opening your notes, then use the longer deck to test the gaps.

A compact core for this board looks like this:

Must not miss coreArtificial Intelligence — Foundations
Supervised learning uses labelled input-target examples; unsupervised learning searches for structure without target labels; reinforcement learning learns from rewards and penalties.
Split data into training, validation and test roles; never use test performance to choose the model or its hyperparameters.
Precision = TP/(TP+FP), recall = TP/(TP+FN), and F1 = 2PR/(P+R); choose metrics that reflect the cost of errors.
Gradient descent updates parameters to reduce a loss; the learning rate controls update size, while regularisation can reduce overfitting.
Transformers use self-attention to build context-aware token representations; a language model commonly predicts the next token.
The must-not-miss core of the artificial intelligence foundations board.

Use the core in two situations. First, recite it at the start of a session to see which area needs attention. Secondly, use it after a full deck review to compress the material again. If you cannot state the core accurately, adding more advanced cards will probably increase familiarity without improving retrieval.

What to do with repeated mistakes

A repeated mistake needs a smaller card, not simply more exposure to the original card. For example, if you repeatedly confuse precision and recall, write a concrete classification scenario and ask which denominator is being used. If you repeatedly forget why a test set must be held out, connect the rule to model selection: using test results to choose between models makes the reported test result part of the development process.

After two failures, place the specific issue in a remediation tray. The prompt should be narrow enough to answer in one or two sentences:

Remediation tray

You lost this card twice: why must the test set remain untouched when selecting a model or tuning hyperparameters?

Add cardDismiss
A remediation prompt created from a repeated error about evaluation data.

Answer the tray card without opening the original answer. A strong response is that the test set is meant to estimate performance on unseen data; if it influences model or hyperparameter selection, information from it has affected development and the estimate is no longer a clean final evaluation. Once you can give that explanation twice on separate reviews, return to the broader validation-set and data-leakage cards.

Build better artificial intelligence flashcards

When adding cards from your own material, preserve the distinctions that are often tested:

  • Ask what precision measures, not “explain classification metrics” in one large card.
  • Ask for the denominator of a metric when formula recall matters.
  • Separate what a transformer does from why attention is useful.
  • Turn a pipeline diagram into cards about the role of each data split.
  • Include a small numerical example when a definition alone is too easy.
  • Add an exception or limitation only when it is supported by your course material.

For instance, a card asking for the definition of data leakage tests recognition. A stronger follow-up asks whether normalising all data before the train-test split can leak information. The answer depends on how the normalisation statistics are calculated: statistics must be fitted on the training data and then applied to validation or test data. That follow-up tests whether you can use the principle, not merely repeat the phrase.

Keep a deck narrow. A board on machine learning evaluation should not also contain every detail of reinforcement learning, computer vision architectures and AI governance. Separate boards make the review queue interpretable and show whether a weakness is in metrics, optimisation, language models or another area.

How MySummaries helps

MySummaries can turn your own artificial intelligence PDFs, slides and handwritten notes into a revision board, then generate focused flashcards, schedule spaced reviews and place repeated errors in a remediation tray. It can also generate written practice and audio explanations from the same board, so the definitions and formulas remain tied to your course material. Start here: MySummaries.