Senior PyTorch Research Engineer

**Job Description:** As a Deep Learning Engineer, your role will involve implementing and benchmarking Liquid Neural Networks (LNNs) for a Doctoral Scholar's thesis on Industrial Edge AI. This project focuses on evaluating the robustness of Neural Circuit Policies (NCPs) and Closed-form Continuous-time (CfC) networks against standard architectures like LSTMs and Transformers using noisy industrial sensor data from NASA CMAPSS. **Key Responsibilities:** - **Data Pipeline Engineering:** - Pre-process the NASA CMAPSS (Turbofan) dataset. - Convert raw sensor logs into sliding-window 3D Tensors (Batch, Time_Steps, Features) suitable for recurrent models. - **Model Implementation (PyTorch):** - Implement a Liquid Neural Network using the ncps (Neural Circuit Policies) library. - Implement baseline models for comparison: Vanilla LSTM and a Lightweight Transformer. - **Robustness Experiments (The Core Task):** - Design a testing loop that injects Gaussian Noise, Signal Dropouts, and Jitter into the test set. - Benchmark inference speed and accuracy degradation across all three models. - **Visualization:** - Generate confusion matrices and RMSE convergence plots. - Use ncps plotting tools to visualize the sparse wiring diagram of the trained Liquid Network. **Qualification Required:** - **Expertise in:** - Python, PyTorch. - **Specific Library Knowledge:** - ncps (by Ramin Hasani/MIT CSAIL) or torchdiffeq. - **Domain Knowledge in:** - Time-Series Anomaly Detection, Recurrent Neural Networks (RNNs). **Bonus Skills:** - Experience with Differential Equations (ODEs) in ML or Edge AI deployment. **Deliverables (Milestone Based):** - **Milestone 1:** - Clean Data Pipeline (NASA CMAPSS) + LSTM Baseline training code. - **Milestone 2:** - Functional Liquid Network (CfC/NCP) training with comparative results. - **Milestone 3:** - Final Robustness Report (Graphs showing Performance vs. Noise Level) + Clean Jupyter Notebooks. **Job Description:** As a Deep Learning Engineer, your role will involve implementing and benchmarking Liquid Neural Networks (LNNs) for a Doctoral Scholar's thesis on Industrial Edge AI. This project focuses on evaluating the robustness of Neural Circuit Policies (NCPs) and Closed-form Continuous-time (CfC) networks against standard architectures like LSTMs and Transformers using noisy industrial sensor data from NASA CMAPSS. **Key Responsibilities:** - **Data Pipeline Engineering:** - Pre-process the NASA CMAPSS (Turbofan) dataset. - Convert raw sensor logs into sliding-window 3D Tensors (Batch, Time_Steps, Features) suitable for recurrent models. - **Model Implementation (PyTorch):** - Implement a Liquid Neural Network using the ncps (Neural Circuit Policies) library. - Implement baseline models for comparison: Vanilla LSTM and a Lightweight Transformer. - **Robustness Experiments (The Core Task):** - Design a testing loop that injects Gaussian Noise, Signal Dropouts, and Jitter into the test set. - Benchmark inference speed and accuracy degradation across all three models. - **Visualization:** - Generate confusion matrices and RMSE convergence plots. - Use ncps plotting tools to visualize the sparse wiring diagram of the trained Liquid Network. **Qualification Required:** - **Expertise in:** - Python, PyTorch. - **Specific Library Knowledge:** - ncps (by Ramin Hasani/MIT CSAIL) or torchdiffeq. - **Domain Knowledge in:** - Time-Series Anomaly Detection, Recurrent Neural Networks (RNNs). **Bonus Skills:** - Experience with Differential Equations (ODEs) in ML or Edge AI deployment. **Deliverables (Milestone Based):** - **Milestone 1:** - Clean Data Pipeline (NASA CMAPSS) + LSTM Baseline training code. - **Milestone 2:** - Functional Liquid Network (CfC/NCP) training with comparative results. - **Milestone 3:** - Final Robustness Report (Graphs showing Performance vs. Noise Level) + Clean Jupyter Notebooks.