Abstract
Technical Project
Most runners don’t realize their mechanics are off until something hurts. Existing wearables can tell you simple performance metrics such as cadence, contact time and stride length, but what they can’t tell you is what’s producing those outputs. How the knee is loading, whether the flexion pattern is asymmetrical or a bad habit is accumulating into a serious injury. That information usually requires a full gait lab. This project aims to use two wearable sensors and a neural network to bridge that gap.
The hardware setup consists of two IMUs, one on the thigh, the other on the calf, each paired with a microcontroller. The microcontrollers are streaming orientation and acceleration data of a walk directly to a computer. That data, along with the user’s physical measurements, are fed into an LSTM model that estimates the knee flexion angle and load on the joint in real-time. The model was trained on a dataset made publicly available by Georgia Tech, consisting of 22 subjects with actual gait lab measurements (goniometers, motion markers, force plates and IMUs) as ground truth. The hardware was designed to match the dataset’s IMU placement as closely as possible so those measurements could be used for training.
While everyday runners can gain helpful insights, the clearest use case is post-ACL rehabilitation. For instance, shortly after ACL Reconstruction Surgery patients are required to train or rehab to break down scar tissue. Having precise insight into the load they put on their knee when doing rehabilitation is a clear objective metric physical therapists could consult. This data would provide more quantified information to physical therapists and a safer and more informed rehabilitation environment.
The system provides real-time estimates of knee flexion angle and joint load. A clinician or user can see exactly how knee load shifts across a session, spot imbalances between legs, or catch high risk patterns without ever going to a gait lab. This isn’t meant to replace clinical assessment, but to make the data that informs it more easily available.
STS Research Project
The paper asks a specific question: If Artificial Neural Networks (ANNs) were historically modeled after the brain, why is their most vital process (learning) modeled after calculus?
The history goes all the way back to McCulloch and Pitts' 1943 logic gates, which framed the neuron as a binary switch, essentially being able to compute any logical function. Then came Rosenblatt's 1958 perceptron, which was still genuinely trying to replicate biological organization. He designed it with the eye and visual cortex in mind, stating “the analogy between the perceptron and biological systems should be readily apparent.” Researchers at the time genuinely thought they were building a synthetic brain.
Minsky and Papert ended that in 1969 by proving that a single-layer perceptron couldn't classify XOR. Funding collapsed and the first AI winter followed. The field came back with Rumelhart, Hinton, and Williams' 1986 backpropagation algorithm, which solved the XOR problem by treating neural networks as continuous differentiable functions propagating error backwards through the entire system. Biological neurons fire in discrete spikes, have no mechanism for backwards weight transport, and don't operate on a global error signal. The gap between the metaphor and the mechanism started in 1986, but the field kept moving, choosing to follow mathematical simplicity instead of biological faithfulness.
Using Thomas Hughes' concept of Technological Momentum, the paper argues that backpropagation didn't just solve a technical problem, it committed the field to a path the surrounding infrastructure made very hard to leave. GPUs happened to excel at the same matrix multiplications backpropagation requires. NVIDIA built its software ecosystem around that. TensorFlow and PyTorch integrated differentiable operations throughout their code. ImageNet made classification accuracy the metric for progress. By the early 2010s, switching back wouldn't have just meant revising the math, it would have meant walking away from the hardware and tooling infrastructure the field stood on.
Sismondo's Discourse Analysis explains why the biological vocabulary ended up outlasting the biological mechanisms. "Neuron," "learning," "neural network" all work as boundary objects; they mean different things to a neurologist than to an engineer, but both can use them without that ambiguity. However, what they then imply becomes the issue. A parameter update called "learning" suggests the system acquired knowledge when it didn’t. That's why a model that hits 99% accuracy on cancer classification can fail at something a first-year medical student wouldn't, and why that failure feels surprising. The word set an expectation the mechanism was never built to meet.
Both projects are asking the same question. The technical project uses a backpropagation-trained LSTM to estimate what a human knee is doing in real time. The STS paper explains why that algorithm has nothing biologically in common with the system it's modeling. The wearable is an instance of the pattern described in the paper: a model of a biological process that's useful if you know what it is and what it does, while misleading if you treat it as something more.
