New GSoC project proposal: Highly Granular Quantization for CICADA (#1668)

* Create proposal_HGQforCICADA.md

* create princeton org (copy from 2023)

* add CICADA project and logo

* Update project proposal

* Update proposal_HGQforCICADA.md

* Update proposal_HGQforCICADA.md

* typo

Co-authored-by: Valentin Volkl <[email protected]>

---------

Co-authored-by: Valentin Volkl <[email protected]>

Author: ligerlac

_gsocorgs/2025/princeton.md

@@ -0,0 +1,11 @@
+---
+title: "Princeton University"
+author: "Lino Gerlach"
+layout: default
+organization: princeton
+logo: princeton-logo.png
+description: |
+  Princeton University is a private Ivy League research university in Princeton, New Jersey.
+---
+
+{% include gsoc_proposal.ext %}

_gsocprojects/2025/project_CICADA.md

@@ -0,0 +1,11 @@
+---
+project: CICADA
+layout: default
+logo: cicada-logo.png
+description: |
+  [CICADA](https://cicada.web.cern.ch/) Calorimeter Image Convolutional Anomaly Detection Algorithm (CICADA) uses low level Compact Muon Solenoid experiment's trigger calorimeter information as the input to convolutional autoencoder to find anomalies produced during Large Hadron Collider proton-proton collisionsis. Quantization Aware Training and Knowledge Distillation are used to compress the model for sub-500ns inference on Field-Programmable Gate Arrays.
+summary: |
+  [CICADA](https://cicada.web.cern.ch/) Calorimeter Image Convolutional Anomaly Detection Algorithm: Real-time anomaly detection at CMS.
+---
+
+{% include gsoc_project.ext %}

_gsocproposals/2025/proposal_HGQforCICADA.md

@@ -0,0 +1,35 @@
+---
+title: Highly Granular Quantization for CICADA
+layout: gsoc_proposal
+project: CICADA
+year: 2025
+organization: princeton
+---
+
+## Description
+
+The CICADA (Calorimeter Image Convolutional Anomaly Detection Algorithm) project aims to provide an unbiased detection of new physics signatures in proton-proton collisions at the Large Hadron Collider's Compact Muon Solenoid experiment (CMS). It detects anomalies in low-level trigger calorimeter information with a  convolutional autoencoder, whose behaviour is transferred to a smaller model through knowledge distillation. Careful quantization of the deployed model allows it to meet the requirement of sub-500ns inference times on FPGAs. While CICADA currently employs Quantization Aware Training with different quantization schemes for each layer of the distilled model, a new gradient-based quantization optimization approach published in 2024 offers the possibility of optimizing quantization at the individual weight level. This project would explore implementing this highly granular quantization method to CICADA's distilled model and evaluating its effects on both model performance and resource consumption on FPGAs. The work would involve implementing the new quantization approach, comparing it with the current implementation, and investigating the impact on both detection performance and hardware resource utilization while maintaining the strict timing requirements.
+
+## Task ideas
+ * Transition CICADA's quantization tool from QKeras to HGQ 
+ * Optimize student model's quantization with higher granularity
+ * Compare resulting model's performance with legacy model
+ * Emulate deployment on FPGA w/ Vivado to evaluate resource consumption  
+
+## Expected results
+ * Extend existing training / quantization scripts to use HGQ in addition to QKeras
+ * A trained student model with highly granular quantization
+ * Estimates of that model's performance and resource consumption on an FPGA
+
+## Requirements
+Python, Tensorflow, Quantization
+
+## Mentors
+  * [Lino Gerlach](mailto:[email protected])
+  * [Isobel Ojalvo](mailto:[email protected])
+  
+## Links
+  * [CICADA (homepage)](https://cicada.web.cern.ch/)
+  * [CICADA (code)](https://github.com/Princeton-AD/cicada)
+  * [HGQ (Paper)](https://arxiv.org/pdf/2405.00645)
+  * [HGQ (code)](https://github.com/calad0i/HGQ)