Fast Machine Learning for Science Conference 2026

Hackathon

Hackathon sessions will take place on Thursday and Friday September 3-4. Please fill out this participation form by August 19 to confirm participation in the hackathon, indicate your project interests, and to receive information on computing resources for use in the hackathon. Teams will be assigned at the start of the hackathon.

Participation modality

The hackathon will be fully in person, with no remote participation option. Remote participants are welcome to attempt the challenges on their own during or after the conference. 

Computing resources

The National Research Platform and San Diego Supercomputer Center are generously providing computing resources for the hackathon, namely a JupyterHub instance with dedicated GPU and FPGA resources, along with container images pre-built for ML and FPGA workflows. Separate datasets for training/testing and final evaluation for all challenges, as well as introductory notebooks for loading the datasets, will be included as part of this JupyterHub. Participants are not required to make use of these resources, but it is highly recommended. 

Timetable

The hackathon will take place during 4 formally organized session blocks: 3 working sessions and 1 presentation/awards session. All formal sessions will take place in The Qualcomm Institute Theater located on the ground floor of Atkinson Hall. Teams are welcome to work together outside of officially programmed sessions. Teams of up to 5 members will be assigned to challenges based on interest and availability during the first hackathon session on Thursday. The hackathon will conclude during the 4th session on Friday from 11:00am - 12:30 pm, where teams will give short summary presentations on their work and judges will choose winning teams and award prizes based on the quality and creativity of their work, their model performance, and their overall presentation. 

Hackathon Challenges

Teams will have a choice of one of four hackathon challenges, including both pre-organized and crowd-sourced challenges: 

• Pre-organized challenges based on the COLLIDE2V (COmprehensive Lhc coLlision Dataset for Foundation model dEvelopment) open-access high energy physics dataset: 
  
  
  • Challenge 1: Di-Higgs (HH→4b vs. QCD) discrimination for trigger systems
    • Task: Binary classification of events as either HH→4b signal or QCD (or other) background.
      
    • Motivation: Identifying boosted Higgs bosons decaying to bottom-quark pairs in the presence of QCD multi-jet background is one of the most important tasks at the HL-LHC and serves as the canonical benchmark for event-tagging algorithms. In particular at the trigger stage, an efficient identification of HH(4b) events is key to recording as many signal events as possible.
    • Dataset: 10 million each HH→4b signal and QCD_HT250toInf background events. 5 million additional background events (TT, W+Jets) (trigger level objects)
    • Evaluation: Area under the ROC curve (AUC) for the network score.

 

• •  Challenge 2: Collider Data Foundation Model Pre-training and Online Adaptation
    • Task: Develop a foundation model, fine-tuning strategy, and associated evaluation metric for a diverse set of physics processes.
    • Motivation: Large-scale pre-training on unlabelled or weakly-labelled collider data followed by task-specific fine-tuning mirrors the paradigm that has driven breakthroughs in natural-language processing and computer vision. This challenge asks whether a similar strategy yields gains in the high-energy physics domain.
    • Dataset: 51 diverse processes with 500k-1million events each. (trigger-level objects) 
    • Evaluation: Performance/quality of model + quality of proposed evaluation metric
•  
  • Challenge 3: Robust Tagging and Self-adapting Trigger Systems Under Changing Detector Conditions
    • Task: Train a model on nominal detector conditions and evaluate its and its ability to adapt as detector degradation occurs.
    • Motivation: During LHC runs, individual detector elements commonly degrade or be temporarily disabled. Trigger and reconstruction algorithms must remain performant under these non-stationary conditions.
    • Dataset: 46 diverse processes with 100k-500k events each. (full-reconstruction objects)
    • Evaluation: AUC versus severity of detector degradation curve. 

 

Many thanks to the COLLIDE2V team, especially Eric Moreno, Maciej Glowacki, Chris Brown, Ellison Scheuller and Arianna Cox for their work in developing these challenges.

• Crowdsourced challenge: As an alternative to the pre-organized challenge track, participants had the option to propose their own hackathon challenges. Many thanks to everyone who submitted challenge proposals!
  
  
  • Challenge 4: Sub-millisecond sEMG Gesture Decoding on FPGA: What Survives Quantization? developed by Siddhant Gulati
    
    • Challenge Description: Prosthetic and brain-computer interface control loops need gesture classification from surface EMG at latencies a user perceives as instantaneous. The published decoders in this space are almost entirely recurrent models running on a host PC over USB, and their reported accuracies are frequently inflated by evaluation protocols that leak information across overlapping windows. This challenge has two halves. First, establish a trustworthy software baseline for multiclass gesture decoding using trial- and session-level splits rather than random window splits. Second, determine which architecture actually wins once FPGA resource and latency constraints are imposed — a quantized BiLSTM, a temporal convolutional network, or a compact transformer — evaluated on the same data under the same budget.
    • Dataset: Primary: emg2pose (Meta), an open-access surface EMG dataset covering roughly 370 hours of recording across 193 participants, released with reference models and evaluation code — . Secondary: Ninapro DB2 and DB5 for smaller-scale cross-dataset checks, freely available for research. Both are open access.
    • Evaluation: Four axes, reported together. (1) Classification accuracy and per-class F1. (2) Inference latency in clock cycles and microseconds. (3) FPGA resource utilization — LUT, DSP, FF, BRAM. (4) Cross-session generalization: train on one session, test on a held-out session. The fourth matters most. Within-session accuracy is the number everyone reports and the one that fails to survive contact with a real user, so a submission that loses a few points of accuracy but holds up across sessions should beat one that does the reverse.

Missing sessions

Birds-of-a-Feather (BoaF) discussion sessions will take place concurrently with hackathon sessions. If you are participating in the hackathon, you will be able to participate in up to one birds-of-a-feather section of your choosing during one of the hackathon session blocks. Please plan to attend all other hackathon sections on Thursday and Friday in order to ensure fair and balanced teams. If you are not participating in the hackathon, you may participate in any or all birds-of-a-feather discussion sessions according to your preference. 

 

Please note:

• If you are planning to attend the first BoaF session Beyond Fast Inference: Toward Adaptive Scientific Systems (Sept 3 11:00am-12:30pm) you will miss the team assignments:
• Please email (mquinnan@ucsd.edu) your preferred challenge (and 2nd most preferred challenge) before Thursday, Sept 3. Otherwise you will be assigned to a team in your absence. 
• If you are planning to attend the last BoaF session LLM Agents for Science: Vibe Coding, Autonomous Scientists, and Beyond (Sept 4 11:00am-12:30pm) you will miss the final presentations, judging, and award ceremony. 
• Please plan to stop by the hackathon room (QI Theater) after the BoaF session to check in with your team and claim any prizes. 
• At least one person from every team will have to be present to give the final presentation, so please keep that in mind when forming your teams.