Različne arhitekture nevronskih omrežij so primerne za različne podatke in naloge — CNN za slike, RNN za sekvence in transformatorji za jezik (in vse več tudi za druge naloge). Razumevanje glavnih vrst pojasni, kako umetna inteligenca rešuje različne probleme.
Glavne arhitekture
CNN (Convolutional Neural Network) → for IMAGES/spatial data:
→ uses convolutions to detect local features (edges, shapes) hierarchically
→ for: image classification, object detection, computer vision
RNN (Recurrent Neural Network) → for SEQUENCES/time-series:
→ processes sequences step by step, maintaining a 'memory' of previous inputs
→ for: text, time-series, speech (older approach; LSTM/GRU variants)
⚠️ struggles with long sequences; largely SUPERSEDED by transformers for language
TRANSFORMER → for SEQUENCES (language) and increasingly everything:
→ attention mechanism; parallel; the dominant modern architecture (LLMs)
→ for: language (LLMs), and now vision, audio, multimodal
