কৃত্রিম বুদ্ধিমত্তা শিখুন - উন্নত স্তর
About this PDF
A valid path need not be the cheapest path. A fluent answer need not have source support. A perfect score may describe only half the inputs. This Advanced book keeps those distinctions visible instead of hiding them behind a single result. All explanations are in Bengali. This 76-page guide runs through 18 units: a task contract before a model score; enumerating a tiny graph before claiming optimality; heuristic quality and the limits of a priority; constraints before preferences; entailment, contradiction and missing facts; minimax on a complete toy game; Bayes updates depend on the population; preprocessing is part of the fitted method; validation splits must match the intended use; a threshold is a decision rule; distance changes when the scale changes; a simple text model and its blind spots; fluency, factual support and task success; retrieval evidence has two stages; abstention changes coverage and the error story; risk controls need evidence and boundaries; monitoring is not a replacement for evaluation; and a capstone comparing a fictional enquiry classifier with a baseline. Every unit gives lesson paragraphs, an original table, a worked case, a controlled paper lab, three questions with explained answers and an extension. An optional Python appendix repeats a few exact synthetic checks with the standard library, and a glossary, study record, acceptance cases and review rubric sit at the back. Every activity uses fictional inputs on paper, with optional harmless offline calculations - no real personal, clinical, financial, employment or confidential records, no account access, no automated action. Invented cases expose errors and support their stated results without proving general algorithm correctness or readiness for real use. Best suited to readers who have finished the Beginner book and want to judge AI claims by evidence, not headlines. Programming is optional; no service account or installation is needed for the paper route. Important: this is a c
Topics covered
Headings from inside this PDF (76 pages in total).
- 8. Preprocessing ও fitted method
- 11. Scale বদলালে distance বদলায়
- 12. এক feature text model এর সীমা
- 14. Retrieval এ দুই ধরনের প্রমাণ
- 15. Abstention এ coverage বদলায়
- 16. Risk control এ প্রমাণ চাই
- 1. A বেশি সঠিক হলেও ঘোষিত output চাহিদা ব্যর্থ করে।
- 2. লাপাত্তা ফল না ভরে table বানান।
- 1. অপরীক্ষিত row কি pass?
- 1. না। প্রমাণের ঘাটতি ও পরের check লিখুন।
What you get
Details
- Format
- PDF · 76 pages
- Pages
- 76 pages
- Language
- Bengali
- Size
- 1.0 MB
- Delivery
- Download to your device, instantly after payment
- Access
- 365 days from purchase
- Sold by
- Sayak Nandy
- Published
- 8 November 2026








