1
NLP & Language Models自然言語処理・言語モデル
Practical NLP work at the level required by research projects — task-specific fine-tuning, embedding-based retrieval, RAG systems, and instruction-tuned model pipelines.
研究プロジェクトで求められる水準の実務的な NLP — タスク特化のファインチューニング、埋め込みベース検索、RAG システム、指示チューニング済みモデルのパイプライン構築 — を対象とします。
Sub-topics and representative methodsサブトピックと代表的な手法
| Understanding & IE理解・情報抽出 |
BERT-family encoders, span-extraction pipelines, NER / relation extraction, structured-output decoding |
| Generation & instruction tuning生成・指示チューニング |
Llama-family / Qwen-family / Mistral-family base and instruct models, SFT, LoRA / QLoRA, DPO |
| Embeddings & retrieval埋め込み・検索 |
BGE, E5, Instructor-style embeddings, ColBERT and late-interaction retrieval, hybrid BM25 + dense |
| RAG systemsRAG |
Chunking strategies, multi-vector retrieval, reranking, evaluation harnesses (Ragas-style), tool-augmented generation |
2
Machine Learning & Deep Learning機械学習・深層学習
General ML/DL work that spans multiple domains — representation learning, self-supervised methods, reinforcement learning, and the underlying training-pipeline engineering.
特定ドメインに収まらない汎用的な ML/DL — 表現学習、自己教師あり学習、強化学習、およびそれらを支える学習パイプラインの設計・実装 — を扱います。
Sub-topics and representative methodsサブトピックと代表的な手法
| Supervised learning教師あり学習 |
Standard tabular and image classification pipelines, calibration, robust training |
| Self-supervised & representation自己教師あり・表現学習 |
Contrastive methods (SimCLR, MoCo), masked-image modeling (MAE), masked-language modeling, multi-modal contrastive |
| Reinforcement learning強化学習 |
PPO, SAC, offline RL (CQL-style), RLHF-style preference pipelines |
| Training craft学習基盤 |
Mixed-precision training, gradient accumulation, deterministic training, distributed data parallel |
3
Graph Algorithms & Recommendationグラフアルゴリズム・推薦
Graph-structured methods, both classical algorithmic and neural, as well as recommendation systems where graph and embedding methods intersect.
古典的なアルゴリズムから現代のニューラル手法まで、グラフ構造を扱う手法全般、およびグラフ・埋め込みが交差する推薦システムを扱います。
Sub-topics and representative methodsサブトピックと代表的な手法
| Graph neural networksグラフニューラルネット |
GCN, GAT, GraphSAGE, message-passing frameworks, heterogeneous-graph models |
| Classical graph algorithms古典グラフアルゴリズム |
Shortest paths, community detection, graph matching, max-flow / min-cut |
| Recommendation systems推薦システム |
Two-tower retrieval, SASRec / BERT4Rec sequential models, PinSAGE-style large-graph retrieval, learning-to-rank |