Perforated LPCVC Track 2 — Dendrites on the Edge (Mobile AI)

Inspiration

The inspiration for this project is that we want to perforate everything. We’re adding dendrites to existing neural networks to squeeze more accuracy out of the same architecture, and this ARM challenge looked like the perfect place to show off where our tech works well. We deliberately wanted to start from a project that was already highly optimized for edge, so that any gains we found couldn't be written off as low-hanging fruit. We believe that if perforation can improve a leader-board topping, hand-tuned edge model, that’s a far better signal than improving an unoptimized one so we started with the winning project for LPCV 2026 Track2. (Challenge details: LPCVC 2026; winners.)

Background

In biology, dendrites are the branching input structures of a neuron, and each one performs its own local, nonlinear computation. This means a single biological neuron behaves more like a small network than the single weighted-sum "unit" that artificial neural networks abstracted it into back in 1943. Perforated Backpropagation restores that missing structure by attaching artificial dendrites to existing network nodes, giving each neuron its own local learning rule so it can extract more accuracy from the same data. The approach builds on Cascade Correlation, a classic technique that grows a network by adding and training new units one at a time to capture the error the existing network cannot. This adds capacity exactly where it's needed instead of scaling everything uniformly. Since perforation improves accuracy at a fraction of traditional scaling, the result is smaller, smarter models that reach target accuracy with far less compute. That efficiency is what lets ML engineers push more capable AI onto constrained edge hardware like ARM chips, with the longer-term aim of loosening the industry's reliance on ever-larger GPU clusters.

What it does

For this hackathon we took the winning LPCVC 2026 Track 2 model (action recognition in video) and perforated it with dendrites. Here's how the perforated model compares to the baseline:

Accuracy. We can see that the Top-1 accuracy improved by just over half a point, from 94.490% to 95.058%. The baseline is already quite strong, so we instead find a more meaningful way to read these results in terms of error reduction. We cut the miss rate from 5.510% to 4.942%, a ~10% relative reduction. On a model that's already been hand-tuned to win a competition, removing 1-in-10 of the remaining errors is strong performance!

Parameters. Perforation added 94,668 parameters, which is about a third of one percent. This is roughly free in terms of footprint for a method, which matters a lot for an edge deployment where memory and model size are constrained. The accuracy gain came at almost no cost to what edge developers care most about.

Throughput. The two models ran at effectively the same speed (2.261 vs. 2.265 clips/s). The difference is well within the run-to-run noise you'd expect on a laptop, so we'd treat inference speed as unchanged rather than make a claim about speedups. The takeaway is that the accuracy improvement did not come with a latency penalty.

In short: meaningfully better accuracy for almost no parameter cost and no measurable speed penalty.

How we built it

We literally just checked out the winning repo (LPCV-Track2-EfficientAI), passed it through our coding agent skill to add perforation, and these were the results we got on our very first try, with zero hyperparameter tuning or experimentation. This is the power of perforation when working with our open source project, also available with an Apache 2.0 license and officially a part of the PyTorch Landscape.

Challenges we ran into

The original training setup runs on Ubuntu with CUDA. To verify ARM compatibility, we had to stand up a PyTorch environment on a Snapdragon laptop running Windows. Windows added some extra friction, but the ARM component itself did not add any complications.

Accomplishments that we're proud of

We always love seeing results like this on the first try. It often takes extra experimentation and parameter sweeps to land gains that are this clean. We're happy our coding assistant tool worked properly and was able to build something working so quickly.

What we learned

• Video classification is a new vertical for us where perforation clearly works.
• Our system runs on ARM.
• The speed impact was surprisingly small. On CUDA our tech typically runs slightly slower; on ARM it was basically identical. One hypothesis: ARM's parallelization may be uniquely compatible with what perforation is doing. Worth investigating further.

What's next for Qualcomm LPCV Challenge Track 2 built for ARM

We plan to write a paper, run hyperparameter sweeps, and add this to our growing repository of example use cases (perforatedai.com/resources), which now includes examples with image classification and segmentation, audio processing, NLP tasks, and many more. Additionally, now that we know these systems are compatible, we pain to do a future experiment applying perforation during pretraining which would allow the creation of a model with matched accuracy but fewer parameters.

Built With

• perforatedai
• pytorch