Awọn awoṣe AI ti ṣalaye adanwo

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

1. Apakan wo ni ohun elo AI jẹ awoṣe ikẹkọ ẹrọ funrararẹ?
   • Awọn igbewọle maapu eto mathematiki si awọn igbejade pẹlu eto ati awọn aye ti ẹkọ
   • Ni wiwo olumulo ti o ni gbogbo ẹya ọja ninu
   • A ṣeto awọn oju opo wẹẹbu awoṣe ti gba laaye lati ṣabẹwo
   • Iwe eto imulo ti o ṣalaye bi ẹgbẹ kan ṣe nlo AI
2. Ẹgbẹ kan ni opin lairi 100 ms iduroṣinṣin. Ninu idanwo alapejuwe itọsọna naa, tani tani pade rẹ?
   • Model B at 94% accuracy and 400 ms
   • Model A at 92% accuracy and 30 ms
   • Neither, because accuracy must be 100%
   • Both, because they score above 90%
3. Ohun elo wo ni hyperparameter dipo paramita ti a kọ ẹkọ?
   • A weight value adjusted during training
   • A coefficient the fitted model uses for an input
   • A tree-depth limit selected through validation
   • A prediction returned for one example
4. Eto wiwa n pese iwe ti a gba pada si awoṣe ni akoko ibeere. Kini iyipada laisi ikẹkọ?
   • It necessarily retrains model weights
   • It fine-tunes the model on every query
   • It permanently changes the model version
   • It adds available evidence without changing model weights
5. Olùgbéejáde itanran-tunse awoṣe kan lori awọn apẹẹrẹ afikun. Kini iyipada?
   • Model weights, without guaranteeing the information is current
   • Only the text retrieved for one query
   • The application’s user interface but never the model
   • The validation set but not model behavior
6. A team adapts a foundation model to a new document task. Which statement about this reuse is accurate?
   • It is necessarily the first AI model ever built
   • It can be adapted to multiple tasks, but still needs evaluation
   • It is only a model for building construction
   • It must be fine-tuned before it can answer any prompt
7. Why should a team re-evaluate after changing a model version?
   • Interfaces always change whenever versions change
   • A new version makes latency measurements unnecessary
   • Behavior can change even while interfaces stay stable
   • A model name proves it is more accurate
8. How should a team choose among candidate models for a defined application?
   • Select the largest parameter count by default
   • Use a leaderboard score without checking its errors
   • Choose the newest version without testing
   • Compare representative inputs, errors and operating constraints, then prefer the simplest adequate option