Ikibazo gihuza ibibazo · Biroroshye Urwego
Uburyo AI Yiga Ikibazo
Reba neza uko wiga imyigire ikurikiranwa, idakurikiranwa kandi ishimangirwa, wongeyeho gukabya, ibizamini, hamwe nigihombo cyikitegererezo gipima.
Ikibazo 1 Bya 8
Mu myigire ikurikiranwa, amakuru y'amahugurwa arimo iki?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Mu myigire ikurikiranwa, amakuru y'amahugurwa arimo iki?
- Only raw examples with no answers attached
- Rules written by programmers for every possible situation
- Inputs paired with target outputs, called labels
- Rewards and penalties from a game environment
- Guteranya abakiriya muburyo busa bwo guhaha nta byiciro byateganijwe mbere ni urugero rwubwoko ki bwo kwiga?
- Kwiga bidakurikiranwa
- Rule-based programming
- Supervised learning
- Kwiga gushimangira
- Nigute umukozi wiga imbaraga, nkumuntu wiga gukina umukino wa videwo, atezimbere?
- By trying actions and learning from rewards or penalties
- By memorizing the game's source code
- By copying labeled examples showing the correct move each turn
- By sorting past games into clusters without feedback
- A model performs well on its training examples but poorly on comparable held-out examples. Which pattern does this show?
- Ntibikwiye
- Kurenza urugero
- Data leakage
- Good generalization
- Why keep a test set separate from fitting the model and choosing its settings?
- To estimate performance on examples not used for those development choices
- To tune hyperparameters by repeatedly checking the final evaluation set
- To add held-out examples to training after the choices are final
- To check whether training predictions match their training labels
- In the guide’s supervised training loop, what does the loss measure?
- How much electricity the training run uses
- How long the model takes to give an answer
- How many examples were deleted from the training dataset
- The numerical difference between predictions and target values
- A text model learns to predict the next token from the preceding context. Which learning setup does the guide describe?
- Supervised learning with separately supplied labels
- Unsupervised clustering of the text
- Self-supervised training that creates a prediction task from the text
- Reinforcement learning from rewards for each token
- A model changes its answer after seeing an example in a prompt. What can be concluded about its learned weights?
- They must have permanently changed
- The prompt became part of every user’s training data
- The model cannot use examples without retraining
- A changed response does not by itself show a weight update