| 研究課題/領域番号 |
23KJ2131
|
|---|
| 研究種目 |
特別研究員奨励費
|
| 配分区分 | 基金 |
|---|
| 応募区分 | 国内 |
|---|
| 審査区分 |
小区分61040:ソフトコンピューティング関連
|
|---|
| 研究機関 | 沖縄科学技術大学院大学 |
|---|
研究代表者 |
Koshkin Roman 沖縄科学技術大学院大学, 科学技術研究科, 特別研究員(DC2)
|
|---|
| 研究期間 (年度) |
2023-04-25 – 2025-03-31
|
| 研究課題ステータス |
交付 (2023年度)
|
| 配分額 *注記 |
1,800千円 (直接経費: 1,800千円)
2024年度: 900千円 (直接経費: 900千円)
2023年度: 900千円 (直接経費: 900千円)
|
|---|
| キーワード | spontaneous activity / spiking neural network / memory / cell assembly / pattern detection |
|---|
| 研究開始時の研究の概要 |
I aim to study spontaneous activity in a biologically realistic memory model (SNN), focusing on the role of cell assemblies for encoding of sequential information (e.g. spatial trajectories), using only a limited number of samples. I will also explore if and how spontaneous activity can serve as a natural mechanism to counter catastrophic forgetting (loss learned information on exposure to new stimuli). I seek to demonstrate that a recurrent network using biologically plausible local learning rules can create durable memory traces, which could have practical applications for the future of AI.
|
| 研究実績の概要 |
Line of work (1) The study of the emergence and properties of structured spontaneous activity in a biologically plausible episodic memory model, with a focus on the role of short-term plasticity and astrocytes for the properties of cell assemblies.
Line of work (2) Development of methods for the detection of patterns in spike data. Published two peer-reviewed papers: (1) Unsupervised Detection of Cell Assemblies with Graph Neural Networks (ICLR 2023) and (2) convSeq: Fast and Scalable Method for Detecting Patterns in Spike Data (ICML 2024).
Line of work (3) Using large language models for simultaneous translation. Status: wrote a paper (posted on arXiv, currently in review at ACL).
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
Line of work (1) Status: The work is mostly finished, I continue to improve the draft of the paper reporting the key findings. Progress on the line of work was slightly slower than expected (the paper should be in review by now). This slight delay is due to the fact that finishing my other work (now published) on the detection of patterns in spike data took significantly more.
Line of work (2) status: Finished, papers published.
Line of work (3) status: Mostly finished, paper in review.
|
| 今後の研究の推進方策 |
I plan to finalize my line of work 1 (The study of the emergence and properties of structured spontaneous activity), and submit the manuscript. Time permitting, I intend to do one more related project (on sequence learning in recurrent networks under biologically realistic plasticity, which will be a logical continuation of the current one).
|
