Biology subjects

Laschowski, B.

Publications and source records attributed to Laschowski, B..

At least 19 recordsLinked to original sources

bioRxiv2026

Interpreting Rewards from Inverse Reinforcement Learning

Inverse reinforcement learning can recover reward functions from observed behavior, but interpreting those rewards remains a fundamental challenge for understanding intelligent behavior and decision-making. To address this challenge, we introduce a novel framework for reward interpretation that combines reward-function analysis, latent mode assignments, and short-history behavioral analysis to infer latent motivations and behavioral dynamics. As a proof-of-concept, we instantiated the framework using switching inverse reinforcement learning on a large-scale dataset of multi-agent social interactions. Our framework interpreted the learned latent modes as cautious and volatile motivational profiles, demonstrating that recovered reward functions can reveal distinct patterns of behavioral dynamics. More broadly, these findings suggest that the proposed framework provides a promising approach for reverse-engineering and interpreting latent rewards underlying intelligent behavior and decision-making.

neuroscience↗

bioRxiv2026

Diffusion Latent Representations for Neural Decoding

Neural decoding can be viewed as a representation learning problem in which neural activity is mapped into an intermediate representation before downstream reconstruction. The choice of intermediate representation influences both performance and learning difficulty. Here we developed a novel framework for studying how intermediate representation choice influences downstream learning and reconstruction. As a proof-of-concept, we instantiated our framework using diffusion latent representations extracted from different diffusion timesteps for neural speech decoding. Component-wise evaluation showed that reconstruction performance differed substantially across diffusion timesteps, with teacher-forced Word Error Rates of 44.7%, 7.5%, and 3.5% for different latent models. These results demonstrate that diffusion latent representations can serve as effective intermediate representations for learning from neural activity, but that their effectiveness depends strongly on the selected diffusion timestep. More broadly, our framework provides a basis for systematically studying how intermediate representation choice influences downstream learning and reconstruction.

neuroscience↗

bioRxiv2026

A brain-inspired framework for memory prioritization in neural networks based on valence

Improving long-term memory in artificial neural networks remains an open challenge. To address this, we developed a novel brain-inspired framework for memory prioritization based on the principle of emotional valence. Our framework includes: (i) a valence-weighted cross-entropy loss that scales the learning signal by the valence magnitude, analogous to neuromodulation; (ii) an amygdala-inspired module that learns high-dimensional valence embeddings; and (iii) a hippocampus-inspired module that integrates valence embeddings into the attention mechanism to modulate information retrieval. We demonstrated the generalization of our framework across spatial, episodic, and language-based memory tasks, consistently improving memory prioritization and long-term retention of high-salience information. In addition to improving long-term memory, we also showed that our framework can help mitigate the "lost-in-the-middle" problem in language modeling. More generally, this research provides further evidence of the potential of brain-inspired algorithms to advance the field of machine learning.

neuroscience↗

bioRxiv2025

An EMG foundation model for neural decoding

Decoding algorithms can be used to predict motor behaviour from patterns of neural activity. However, most studies rely on subject-optimized models, limiting generalization and scalability to novel subjects and tasks. Building on recent advances in deep learning and large-scale data, here we developed an EMG foundation model for neural decoding. Our model was trained on more than 197 hours of neural recordings from 1,667 subjects. We used unsupervised learning to pretrain our encoder layers on unlabeled data, followed by supervised learning on our benchmark dataset. Additionally, we performed large-scale architecture searches to develop a custom encoder-decoder model composed of convolutional and transformer layers, optimized for both scalability and performance. Our foundation model consistently outperformed the previous state-of-the-art (i.e., subject-optimized models) across both in-distribution and out-of-distribution evaluations. For in-distribution evaluation, few-shot fine-tuning yielded an average F1 score of 0.697, compared to 0.638 for subject-optimized models. For out-of-distribution evaluation on clinical and demographically-shifted subjects, we achieved an average F1 score of 0.599, compared to 0.518 for the subject-optimized baselines. Taken together, our results highlight the value of foundation models for robust and generalizable neural decoding. By publicly releasing our neural network weights and training pipeline, we aim to support future research in computational neuroscience and neural-machine interfaces.

neuroscience↗

bioRxiv2025

Mixture models for domain-adaptive brain decoding

A grand challenge in brain decoding is to develop algorithms that generalize across multiple subjects and tasks. Here, we developed a new computational framework to minimize negative transfer for domain-adaptive brain decoding by reframing source selection as a mixture model parameter estimation problem, allowing each source subject to contribute through a continuous mixture weight rather than being outright included or excluded. To compute these weights, we developed a novel convex optimization algorithm based on the Generalized Method of Moments. By using model performance metrics as the generalized moment functions in our GMM optimization, our algorithm also provides theoretical guarantees that the mixture weights are an optimal approximation of the importance weights that underlie domain adaptation theory. When tested on a large-scale brain decoding dataset (n=105 subjects), our new mixture model weighting framework achieved state-of-the-art performance--increasing accuracy up to 2.5% over baseline fine-tuning, double the performance gain compared to previous research in supervised source selection. Notably, these improvements were achieved using significantly less training data (i.e., 62% smaller effective sample sizes), suggesting that our performance gains stem from reduced negative transfer. Collectively, this research advances toward a more principled and generalizable brain decoding framework, laying the mathematical foundation for scalable brain-computer interfaces and other applications in computational neuroscience.

neuroscience↗

bioRxiv2024

Development of a real-time neural controller using an EMG-driven musculoskeletal model

Here we present our development of a novel real-time neural controller based on an EMG-driven musculoskeletal model, designed for volitional control of robots and computers. Our controller uniquely enables motion control during both isometric and non-isometric muscle contractions. We address several key challenges in EMG control system design, including accuracy, latency, and robustness. Our approach combines EMG signal processing, neural activation dynamics, and Hill-type muscle modeling to translate neural commands into muscle forces, which can enhance robustness against electrode variability and signal noise. Additionally, we integrate muscle activation dynamics with impedance control, inspired by the human motor control system, for smooth and adaptive interactions. As an initial proof of concept, we demonstrated that our system could control a robot actuator across a range of movements, both static and dynamic, and at different operating speeds, achieving high reference tracking performance and state-of-the-art processing times of 2.9 ms, important for real-time embedded computing. This research helps lay the groundwork for next-generation neural-machine interfaces that are fast, accurate, and adaptable to diverse users and control applications.

neuroscience↗

bioRxiv2024

Comparative analysis of neural decoding algorithms for brain-machine interfaces

Accurate neural decoding of brain dynamics remains a significant and open challenge in brain-machine interfaces. While various signal processing, feature extraction, and classification algorithms have been proposed, a systematic comparison of these is lacking. Accordingly, here we conducted one of the largest comparative studies evaluating different combinations of state-of-the-art algorithms for motor neural decoding to find the optimal combination. We studied three signal processing methods (i.e., artifact subspace reconstruction, surface Laplacian filtering, and data normalization), four feature extractors (i.e., common spatial patterns, independent component analysis, short-time Fourier transform, and no feature extraction), and four machine learning classifiers (i.e., support vector machine, linear discriminant analysis, convolutional neural networks, and long short-term memory networks). Using a large-scale EEG dataset, we optimized each combination for individual subjects (i.e., resulting in 672 total experiments) and evaluated performance based on classification accuracy. We also compared the computational and memory storage requirements, which are important for real-time embedded computing. Our comparative analysis provides novel insights that help inform the design of next-generation neural decoding algorithms for brain-machine interfaces used to interact with and control robots and computers.

neuroscience↗

bioRxiv2024

Egocentric perception of walking environments using an interactive vision-language system

Large language models can provide a more detailed contextual understanding of a scene beyond what computer vision alone can provide, which have implications for robotics and embodied intelligence. In this study, we developed a novel multimodal vision-language system for egocentric visual perception, with an initial focus on real-world walking environments. We trained a number of state-of-the-art transformer-based vision-language models that use causal language modelling on our custom dataset of 43,055 image-text pairs for few-shot image captioning. We then designed a new speech synthesis model and a user interface to convert the generated image captions into speech for audio feedback to users. Our system also uniquely allows for feedforward user prompts to personalize the generated image captions. Our system is able to generate detailed captions with an average length of 10 words while achieving a high ROUGE-L score of 43.9% and a low word error rate of 28.1% with an end-to-end processing time of 2.2 seconds. Overall, our new multimodal vision-language system can generate accurate and detailed descriptions of natural scenes, which can be further augmented by user prompts. This innovative feature allows our image captions to be personalized to the individual and immediate needs and preferences of the user, thus optimizing the closed-loop interactions between the human and generative AI models for understanding and navigating of real-world environments.

neuroscience↗

bioRxiv2023

Reinforcement Learning for Control of Human Locomotion in Simulation

Control of robotic leg prostheses and exoskeletons is an open challenge. Computer modeling and simulation can be used to study the dynamics and control of human walking and extract principles that can be programmed into robotic legs to behave similar to biological legs. In this study, we present the development of an efficient two-layer Q-learning algorithm, with k-d trees, that operates over continuous action spaces and a reward model that estimates the degree of muscle activation similarity between the agent and human state-to-action pairs and state-to-action sequences. We used a human musculoskeletal model acting in a high-dimensional, physics-based simulation environment to train and evaluate our algorithm to simulate biomimetic walking. We used imitation learning and artificial bio-mechanics data to accelerate training via expert demonstrations and used experimental human data to compare and validate our predictive simulations, achieving 79% accuracy. Also, when compared to the previous state-of-the-art that used deep deterministic policy gradient, our algorithm was significantly more efficient with lower computational and memory storage requirements (i.e., requiring 7 times less RAM and 87 times less CPU compute), which can benefit real-time embedded computing. Overall, our new two-layer Q-learning algorithm using sequential data for continuous imitation of human locomotion serves as a first step towards the development of bioinspired controllers for robotic prosthetic legs and exoskeletons. Future work will focus on improving the prediction accuracy compared to experimental data and expanding our simulations to other locomotor activities.

bioengineering↗

bioRxiv2023

Sequential image classification of human-robot walking environments using temporal neural networks

Robotic prosthetic legs and exoskeletons require real-time and accurate estimation of the walking environment for smooth transitions between different locomotion mode controllers. However, previous studies have mainly been limited to static image classification, therein ignoring the temporal dynamics of human-robot locomotion. Motivated by these limitations, here we developed several state-of-the-art temporal convolutional neural networks (CNNs) to compare the performances between static vs. sequential image classification of real-world walking environments (i.e., level-ground terrain, incline stairs, and transitions to and from stairs). Using our large-scale image dataset, we trained a number of encoder networks such as VGG, MobileNetV2, ViT, and MobileViT, each coupled with a temporal long short-term memory (LSTM) backbone. We also trained MoViNet, a new video classification model designed for mobile and embedded devices, to further compare the performances between 2D and 3D temporal deep learning models. Our 3D network outperformed all the hybrid 2D encoders with LSTM backbones and the 2D CNN baseline model in terms of classification accuracy, suggesting that network architecture can play an important role in performance. However, although our 3D neural network achieved the highest classification accuracy, it had disproportionally higher computational and memory storage requirements, which can be disadvantageous for real-time control of robotic leg prostheses and exoskeletons with limited onboard resources.

bioengineering↗

bioRxiv2023

AI-powered smart glasses for sensing and recognition of human-robot walking environments

Environment sensing and recognition can allow humans and robots to dynamically adapt to different walking terrains. However, fast and accurate visual perception is challenging, especially on embedded devices with limited computational resources. The purpose of this study was to develop a novel pair of AI-powered smart glasses for onboard sensing and recognition of human-robot walking environments with high accuracy and low latency. We used a Raspberry Pi Pico microcontroller and an ArduCam HM0360 low-power camera, both of which interface with the eyeglass frames using 3D-printed mounts that we custom-designed. We trained and optimized a lightweight and efficient convolutional neural network using a MobileNetV1 backbone to classify the walking terrain as either indoor surfaces, outdoor surfaces (grass and dirt), or outdoor surfaces (paved) using over 62,500 egocentric images that we adapted and manually labelled from the Meta Ego4D dataset. We then compiled and deployed our deep learning model using TensorFlow Lite Micro and post-training quantization to create a minimized byte array model of size 0.31MB. Our system was able to accurately predict complex walking environments with 93.6% classification accuracy and had an embedded inference speed of 1.5 seconds during online experiments using the integrated camera and microcontroller. Our AI-powered smart glasses open new opportunities for visual perception of human-robot walking environments where embedded inference and a low form factor is required. Future research will focus on improving the onboard inference speed and miniaturization of the mechatronic components.

bioengineering↗

bioRxiv2023

Efficient Visual Perception of Human-Robot Walking Environments Using Semi-Supervised Learning

Convolutional neural networks trained using supervised learning can improve visual perception for human-robot walking. These advances have been possible due to large-scale datasets like ExoNet and StairNet - the largest open-source image datasets of real-world walking environments. However, these datasets require vast amounts of manually annotated data, the development of which is time consuming and labor intensive. Here we present a novel semi-supervised learning system (ExoNet-SSL) that uses over 1.2 million unlabelled images from ExoNet to improve training efficiency. We developed a deep learning model based on mobile vision transformers and trained the model using semi-supervised learning for image classification. Compared to standard supervised learning (98.4%), our ExoNet-SSL system was able to maintain high prediction accuracy (98.8%) when tested on previously unseen environments, while requiring 35% fewer labelled images during training. These results show that semi-supervised learning can improve training efficiency by leveraging large amounts of unlabelled data and minimize the size requirements for manually annotated images. Future research will focus on model deployment for onboard real-time inference and control of human-robot walking.

bioengineering↗

bioRxiv2023

Preliminary Development of a Robotic Hip-Knee Exoskeleton with 3D-Printed Backdrivable Actuators

Robotic exoskeletons can provide powered locomotor assistance and rehabilitation to persons with mobility impairments due to aging and/or physical disabilities. Here we present the preliminary development and systems integration of T-BLUE - a modular, bilateral robotic hip-knee exoskeleton with 3D-printed backdriveable actuators. We retrofitted commercially available passive postoperative orthoses with open-source 3D-printed actuators to minimize cost and improve accessibility. The actuators are of quasi-direct drive design with high-torque density brushless DC motors and low gearing (15:1 transmission ratio) for low output impedance and high backdrivability, therein allowing for energy-efficient and dynamic human-robot physical interaction and legged locomotion. The modular design allows the exoskeleton to be customized and adapted to different users (e.g., persons with lateral vs. bilateral mobility impairments) and different hip-knee joint configurations. The goals of this preliminary study were to describe our experience with regards to the repeatability of the open-source 3D-printed actuators in engineering practice and the feasibility of integrating the actuators into wearable robotics hardware. This qualitative research serves as a first step towards using the robotic exoskeleton to support the development and testing of novel controller designs and rehabilitation protocols for different locomotor activities of daily living. We are especially interested in populations that could benefit from partial locomotor assistance such as older adults and/or persons with osteoarthritis. Future research will involve benchtop testing to quantitatively evaluate the actuator performance in terms of dynamics and energy-efficiency.

bioengineering↗

bioRxiv2023

Development and Mobile Deployment of a Stair Recognition System for Human-Robot Locomotion

Environment sensing and recognition can improve the safety and autonomy of human-robot locomotion, especially during transitions between environmental states such as walking to and from stairs. However, accurate and real-time perception on edge devices with limited computational resources is an open problem. Here we present the development and mobile deployment of StairNet - a vision-based automated stair recognition system powered by deep learning. Building on ExoNet - the largest open-source dataset of egocentric images of real-world walking environments - we designed a new dataset specifically for stair recognition with over 515,000 images. We then developed a lightweight and efficient convolutional neural network for image classification, which accurately predicted complex stair environments with 98.4% accuracy. We also studied different model compression and optimization methods and deployed our system on several mobile devices running a custom-designed iOS application with onboard accelerators using CPU, GPU, and/or NPU backend computing. Of the designs that we tested, our highest performing system showed negligible reductions in classification accuracy due to the model conversion for mobile deployment and achieved an inference time of 2.75 ms on an iPhone 11. The high speed and accuracy of the StairNet system on edge devices opens new opportunities for autonomous control and planning of robotic prosthetic legs, exoskeletons, and other assistive technologies for human locomotion.

bioengineering↗

bioRxiv2023

Continuous Prediction of Leg Kinematics during Walking using Inertial Sensors, Smart Glasses, and Embedded Computing

Unlike traditional hierarchical controllers for robotic leg prostheses and exoskeletons, continuous systems could allow persons with mobility impairments to walk more naturally in real-world environments without requiring high-level switching between locomotion modes. To support these next-generation controllers, we developed a new system called KIFNet (Kinematics and Image Fusing Network) that uses lightweight and efficient deep learning models to continuously predict the leg kinematics during walking. We tested different sensor fusion methods to combine kinematics data from inertial sensors and computer vision data from smart glasses and found that adaptive instance normalization achieved the lowest RMSE predictions for knee and ankle joint kinematics. We also deployed our model on an embedded device. Without inference optimization, our model was 20 times faster than the previous state-of-the-art and achieved 20% higher prediction accuracies, and during some locomotor activities like stair descent, decreased RMSE up to 300%. With inference optimization, our best model achieved 125 FPS on an NVIDIA Jetson Nano. These results demonstrate the potential to build fast and accurate deep learning models for continuous prediction of leg kinematics during walking based on sensor fusion and embedded computing, therein providing a foundation for real-time continuous controllers for robotic leg prostheses and exoskeletons.

bioengineering↗

bioRxiv2022

Simulation of Energy Regeneration in Human Locomotion for Efficient Exoskeleton Actuation

Backdriveable actuators with energy regeneration can improve the efficiency and extend the battery-powered operating times of robotic lower-limb exoskeletons by converting some of the otherwise dissipated energy during negative mechanical work into electrical energy. However, previous related studies have focused on steady-state level-ground walking. To better encompass real-world community mobility, here we developed a feedforward human-exoskeleton energy regeneration system model to simulate energy regeneration and storage during other daily locomotor activities. Data from inverse dynamics analyses of 10 healthy young adults walking at variable speeds and slopes were used to calculate the negative joint mechanical power and work (i.e., the mechanical energy theoretically available for electrical energy regeneration). These human joint mechanical energetics were then used to simulate backdriving a robotic exoskeleton and regenerating energy. An empirical characterization of the exoskeleton device was carried out using a joint dynamometer system and an electromechanical motor model to calculate the actuator efficiency and to simulate energy regeneration. Our performance calculations showed that regenerating energy at slower walking speeds and decline slopes could significantly extend the battery-powered operating times of robotic lower-limb exoskeletons (i.e., up to 99% increase in total number of steps), therein improving locomotor efficiency.

bioengineering↗

bioRxiv2022

Stair Recognition for Robotic Exoskeleton Control using Computer Vision and Deep Learning

Computer vision can be used in robotic exoskeleton control to improve transitions between different locomotion modes through the prediction of future environmental states. Here we present the development of a large-scale automated stair recognition system powered by convolutional neural networks to recognize indoor and outdoor real-world stair environments. Building on the ExoNet database - the largest and most diverse open-source dataset of wearable camera images of walking environments - we designed a new computer vision dataset, called StairNet, specifically for stair recognition with over 515,000 images. We then developed and optimized an efficient deep learning model for automatic feature engineering and image classification. Our system was able to accurately predict complex stair environments with 98.4% classification accuracy. These promising results present an opportunity to increase the autonomy and safety of human-exoskeleton locomotion for real-world community mobility. Future work will explore the mobile deployment of our automated stair recognition system for onboard real-time inference.

bioengineering↗

bioRxiv2021

Environment Classification for Robotic Leg Prostheses and Exoskeletons using Deep Convolutional Neural Networks

Robotic leg prostheses and exoskeletons can provide powered locomotor assistance to older adults and/or persons with physical disabilities. However, the current locomotion mode recognition systems being developed for intelligent high-level control and decision-making use mechanical, inertial, and/or neuromuscular data, which inherently have limited prediction horizons (i.e., analogous to walking blindfolded). Inspired by the human vision-locomotor control system, we designed and evaluated an advanced environment classification system that uses computer vision and deep learning to forward predict the oncoming walking environments prior to physical interaction, therein allowing for more accurate and robust locomotion mode transitions. In this study, we first reviewed the development of the ExoNet database - the largest and most diverse open-source dataset of wearable camera images of indoor and outdoor real-world walking environments, which were annotated using a hierarchical labelling architecture. We then trained and tested over a dozen state-of-the-art deep convolutional neural networks (CNNs) on the ExoNet database for large-scale image classification of the walking environments, including: EfficientNetB0, InceptionV3, MobileNet, MobileNetV2, VGG16, VGG19, Xception, ResNet50, ResNet101, ResNet152, DenseNet121, DenseNet169, and DenseNet201. Lastly, we quantitatively compared the benchmarked CNN architectures and their environment classification predictions using an operational metric called NetScore, which balances the image classification accuracy with the computational and memory storage requirements (i.e., important for onboard real-time inference). Although we designed this environment classification system to support the development of next-generation environment-adaptive locomotor control systems for robotic prostheses and exoskeletons, applications could extend to humanoids, autonomous legged robots, powered wheelchairs, and assistive devices for persons with visual impairments.

bioengineering↗