Research notes · Experiments · Tutorials
Notes on reinforcement learning, energy, research methods, and the systems I am building. Browse by series or jump into any article below.
Series 1
Revisiting our 2019 study of physiological pain recognition: why different feature sets can encode similar information, and what a topological map adds to an accuracy score.
Read article →Revisiting feature-space differences between people with and without limb loss: why a feature set’s complementarity may change when it reaches a different population.
Read article →Revisiting Bayesian costs and priors in LDA: how changing the decision rule can reduce active errors and unstable switching without improving overall accuracy.
Read article →Revisiting our EMG-processing review: how time, noise, force, position, and different users connect signal-processing choices to the conditions an interface must handle.
Read article →Revisiting our comparison of myoelectric patterns across 20 intact-limbed participants and ten participants with limb loss, using clustering, principal components, and classification to characterize inter-subject variability.
Read article →A review and six-dataset analysis quantify two sources of EMG variation and test how training coverage changes their effects.
Read article →Two public datasets help explain how position constraints change the information available to gesture recognizers.
Read article →A feasibility study tests the fidelity of generated muscle signals and their value as classifier-training augmentation.
Read article →Handcrafted features provide reference points for interpreting a deep EMG model’s learned representations.
Read article →A cross-user EMG model adapts to a new participant using a small calibration set, with separate results for intact limbs and limb loss.
Read article →A pilot cursor-control study compares positive and negative task context with confidence-based model updates.
Read article →The original library paper connects data handling, signal processing, recognition, and interactive evaluation in one modular toolkit.
Read article →EMaGer’s uniform electrode geometry enables a specific augmentation method for sensor shifts and between-session recognition.
Read article →A large public EMG dataset makes it possible to measure cross-user recognition, training-set size, and the limits of generalization.
Read article →The BioPoint study tests a compact EMG, inertial, and optical sensing system for sustained gesture recognition and rhythm-game control.
Read article →Two experiments test context-informed learning with sparse calibration and after a deliberate 45-degree armband rotation.
Read article →A twelve-person study compares EMG, acceleration, and PPG for event recognition across repetitions, positions, and days.
Read article →A small exploratory study examines anticipation and all-or-nothing contractions when collecting regression training data.
Read article →A multimodal transfer-learning study estimates upper-limb force across users and contraction conditions.
Read article →Contrastive training uses examples of daily activity to improve rejection while preserving recognition of the target gestures.
Read article →The MEC24 report examines context-informed updates after short calibration and after electrode displacement.
Read article →The original on-demand-control preprint tests finger-snap mode switching during everyday activity.
Read article →Context-informed training is compared with prompted calibration, using online throughput and recorded-gesture accuracy as distinct outcomes.
Read article →A controlled study quantifies three confounds for complete-gesture recognition and tests representative-data aggregation.
Read article →A narrative review separates the information used for adaptation from the kind of change the controller is meant to learn.
Read article →A nine-person VR study uses object-specific grasp context to update recognition during a functional task.
Read article →An equal-training-time comparison examines throughput, drift, and unwanted simultaneous activation in two-degree-of-freedom control.
Read article →The journal wake-gesture study evaluates unintended toggles, missed snaps, and user experience during alarm and robot tasks.
Read article →LibEMG’s dataset extension standardizes access and evaluation across twenty-one datasets while preserving their different tasks and conditions.
Read article →A twelve-person study examines signal changes, recognition across thermal conditions, and temperature-diverse training.
Read article →Public data and released models support new-user cursor and discrete-gesture tasks, with explicit practice and task limits.
Read article →TFTune learns gain mappings for a trackpad, Windows mice, and an EMG cursor, then evaluates the resulting speed–accuracy tradeoffs.
Read article →A transformer with context-informed adaptation and soft dynamic time warping is evaluated across three target precision levels.
Read article →Series 2
Build a reproducible gesture-recognition experiment from the public Myo dataset, preserving recording boundaries from loading through evaluation.
Read article →Inspect EMG in time and frequency, then use controlled motion-like drift and power-line interference to understand what bandpass and notch filters remove.
Read article →Derive two model-based estimators and test their assumptions: Wiener filtering for a noisy waveform and Kalman filtering for a changing amplitude state.
Read article →Connect ten feature families to their equations, compare them with the same LDA classifier, and examine redundancy within and between gestures.
Read article →Use grouped forward selection and an explicit accuracy tolerance to choose a small feature subset before evaluating it on held-out recordings.
Read article →Train a small feature MLP and a temporal convolutional network with the same recording split, then compare learning curves and what each model is asked to learn.
Read article →Pair EMG and orientation recordings from 3DC, build a trial-level fusion baseline, and test whether the second modality contributes useful information.
Read article →Use LibEMG’s offline data and filtering tools with WFDB ECG recordings, then build and evaluate an educational QRS and R-peak timing pipeline.
Read article →Series 3
A bandit problem models sequential decision-making under uncertainty, where every step an agent selects one action (an arm), receives a scalar reward, and updates its belief about that arm.
Read article →The Bellman equation is a fundamental recursive relationship in dynamic programming and reinforcement learning that expresses the value of a decision problem at a certain point in time in terms of the value at subsequent points in time.
Read article →Dynamic programming (DP) is a collection of algorithms that can be used to solve reinforcement learning problems when a perfect model of the environment is known. DP methods break down problems into smaller subproblems and solve them recursively.
Read article →Monte Carlo methods are a class of algorithms that rely on repeated random sampling to obtain numerical results. In reinforcement learning, they are used to estimate the value of states or actions based on the average return observed from multiple episodes.
Read article →Temporal-Difference (TD) learning is a reinforcement learning approach tries to improve esimates based in part on other learned estimates, without waiting for a final outcome (as in Monte Carlo methods).
Read article →Replacing a table with a learned function: neurons, gradient descent, backpropagation, and the bridge to deep reinforcement learning.
Read article →Turning the Bellman update into a neural-network loss, then separating experience replay, target networks, Double DQN, and dueling architectures.
Read article →Using complete trajectories and log probabilities to increase the likelihood of actions that lead to better returns.
Read article →Learning a critic to guide policy updates, then extending the idea to continuous actions with deterministic policy gradients.
Read article →Twin critics, delayed policy updates, target smoothing, and entropy-regularized control—what each mechanism changes.
Read article →
