Merge pull request #284 from neural-loop/student-talk

Student talk

Author: neural-loop

.github/workflows/main.yml

@@ -2,7 +2,7 @@ name: Deploy Hugo site to Pages
 
 on:
   push:
-    branches: ["main", "cache-images"]
+    branches: ["main", "student-talk"]
 
 permissions:
   contents: read
@@ -63,7 +63,7 @@ jobs:
 
       - name: Generate OG Images
         run: npm run og-images
-        
+
       - name: Fetch GitHub Stars
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.gitignore

@@ -26,3 +26,4 @@ yarn.lock
 content/**/*-og-*.jpg
 static/images/og-image-*.jpg
 tmp/og-cache-manifest.json
+/data/github_stars.json

content/contributors/matei-stan/index.md

@@ -0,0 +1,16 @@
+---
+title: "Matei Stan"
+description: "Matei Stan is a PhD student at the University of Manchester, focusing on State Space Models (SSMs) and their application in Spiking Neural Networks for long-range sequential tasks."
+image: "matei-stan.jpg"
+social:
+  - icon: "fa-brands fa-linkedin"
+    link: "https://www.linkedin.com/in/matei-stan"
+    title: "linkedin"
+draft: false
+---
+
+Matei Stan is a third-year PhD student in the Department of Computer Science at the University of
+Manchester, UK. He is supervised by Dr Oliver Rhodes in the Advanced Processor Technologies
+(APT) group. In his PhD work, Matei has primarily focused on the applications of deep State Space
+Models (SSMs), such as S4, in neuromorphic computing, and their potential in scaling
+energy-efficient algorithms for long-range sequential tasks.

content/contributors/matei-stan/matei-stan.jpg

@@ -0,0 +1,31 @@
+---
+title: "Student Talk: Learning Long Sequences in Spiking Neural Networks with Matei Stan"
+author:
+  - "Matei Stan"
+date: 2025-07-27
+start_time: "08:30"
+end_time: "09:45"
+time_zone: "EST"
+description: "Explore how State Space Models (SSMs) combined with Spiking Neural Networks (SNNs) can outperform Transformers on long-sequence tasks, and learn about a novel feature mixing layer that challenges assumptions about binary activations."
+upcoming: true
+upcoming_url: "https://teams.microsoft.com/l/meetup-join/19%3Ameeting_OTBkNTY5MjgtMjE3Ni00OTFmLWEwNzktN2QwZTU1NWIxNDc2%40thread.v2/0?context=%7B%22Tid%22%3A%22c152cb07-614e-4abb-818a-f035cfa91a77%22%2C%22Oid%22%3A%223f444780-d657-4917-993e-0f42adeff90e%22%7D"
+video: ""
+image: "banner.png"
+speaker_photo: "matei-stan.jpg"
+type: "student-talks"
+speaker_bio: "Matei Stan is a third-year PhD student in the Department of Computer Science at the University of Manchester, UK. He is supervised by Dr Oliver Rhodes in the Advanced Processor Technologies (APT) group. In his PhD work, Matei has primarily focused on the applications of deep State Space Models (SSMs), such as S4, in neuromorphic computing, and their potential in scaling energy-efficient algorithms for long-range sequential tasks."
+---
+
+Matei’s published work, “Learning Long Sequences in Spiking Neural Networks” [1], systematically investigates, for the first time, the intersection of the State‑of‑The‑Art State Space Models (SSMs) with Spiking Neural Networks (SNNs) for long‑range sequence modelling. Results suggest that SSM‑based SNNs can outperform the Transformer on all tasks of a well‑established long‑range sequence modelling benchmark - the “Long-Range Arena” [2]. It is also shown that the SSM‑based SNNs can outperform current State‑of‑The‑Art SNNs with fewer parameters on sequential image classification. Finally, a novel feature mixing layer is introduced, improving SNN accuracy while challenging assumptions about the role of binary activations in SNNs. This work paves the way for deploying powerful SSM-based architectures, such as Large Language Models, on neuromorphic hardware for energy-efficient long-range sequence modelling.
+
+This talk will highlight, at a high level, the similarities in computational primitives between SSMs and the existing neuromorphic standards such as Leaky Integrate-and-Fire (LIF) neurons. It will also focus on the specific drawbacks brought about by the introduction of binary activations in SSMs, as well as the extent to which these can be mitigated by the development of more accurate surrogate gradient methods that account for non-differentiability. Finally, arguments will be presented in favour of separating biological plausibility from energy efficiency in attempting to create scalable neuromorphic solutions.
+
+**Contents of the talk:**
+- Introduction to long-range sequence modelling
+- Link between SSMs and SNNs
+- Binary S4D and the Gated Spiking Unit
+- Q&A
+
+**References**
+[1]: Stan, M.I. and Rhodes, O., 2024. Learning long sequences in spiking neural networks. Scientific Reports, 14(1), p.21957.
+[2]: Tay, Y., Dehghani, M., Abnar, S., Shen, Y., Bahri, D., Pham, P., Rao, J., Yang, L., Ruder, S. and Metzler, D., 2020. Long range arena: A benchmark for efficient transformers. arXiv preprint arXiv:2011.04006.