I work on recommender systems, applying LLMs and RL methods.

I prepared half of the practical coding materials and exercises, 6 out of 12 Python notebooks, for the "Reinforcement Learning" graduate course. Topics I covered:1. Dynamic Programming: Value Iteration and Policy Iteration for the car-sharing problem;2. Linear-Quadratic Regulator (LQR): Linearly constrained least-squares and Dynamic Programming solutions of the moving vehicle problem;3. Imitation Learning: Learning the Ant robot to walk in MuJoCo using Dagger;4. Soft Actor-Critic: Robust pendulum stabilization;5. Monte-Carlo Tree Search: Making super-human Connect4 player;6. Model-based Reinforcement Learning: Learning expressive world models and training to play VizDoom Take Cover in dreams;Moreover, I taught the practical class, for the group of 10 or more graduate students, covering the above and the remaining topics: Bandit algorithms; Monte-Carlo Control; Q-Learning; SARSA; Deep Q-Networks; Actor-critic methods for continuous control; Offline RL; Meta-RL; Multi-task RL;In my teaching, I emphasize intuitive understanding through e.g. rich visualization.

Projects I worked on: - Bayesian methods for Deep Reinforcement Learning: Januszewski, Piotr, Mateusz Olko, Michał Królikowski, Jakub Świątkowski, Marcin Andrychowicz, Łukasz Kuciński, and Piotr Miłoś. ‘Continuous Control With Ensemble Deep Deterministic Policy Gradients’. In 2021 NeurIPS DeepRL Workshop, 2021. https://openreview.net/forum?id=TIUfoXsnxB. - Model-based Reinforcement Learning in Google Research Football: Czechowski, Konrad, Piotr Januszewski, Piotr Kozakowski, Łukasz Kuciński, and Piotr Miłoś. ‘Structure and Randomness in Planning and Reinforcement Learning’. In 2021 International Joint Conference on Neural Networks (IJCNN), 2021. https://doi.org/10.1109/IJCNN52387.2021.9533317. - Development and maintenance of distributed Python frameworks for model-based RL (partially published here: https://github.com/subgoal-search/subgoal-search/tree/main/alpacka) and model-free RL (https://github.com/awarelab/spinningup_tf2).

I've worked on the "Toddler-Inspired Active Representation Learning" project. Under the supervision of prof. João F. Henriques and prof. Weidi Xie from the Visual Geometry Group (VGG), I've trained reinforcement learning (RL) PPO agents in the meta-RL loop with aim to optimize the visual data collection process for the semi-supervised representation learning SimCLR method. Beyond theoretical results, it concluded with the efficient distributed meta-RL training code written in PyTorch: https://github.com/piojanu/active-representation-learning

I worked as Machine Learning and Deep Learning trainer at InfoShare Academy. My classes put an emphasis on the practical aspects of ML and DL. Moreover, I advice students in their own projects.Topic I taught so far:- Linear regression, confidence intervals, hypothesis testing, - Introduction to Artificial Neural Networks and Deep Learning,- Computer vision: classification, object detection and semantic segmentation using Deep Learning,- Training and deployment of deep models using OpenVINO: "Object detection from cloud to the edge",- Deep Reinforcement Learning on the example of AlphaZero.

I've written 10 blog posts for the Neptune.ai blog on topics related to machine/deep/reinforcement learning, software engineering, productivity, and team management. Available here: https://neptune.ai/blog/author/piotrjanuszewski

The Intel® Software Innovator program supports innovative developers who display an ability to create forward-looking projects like mine. I’ve been able to i.a. share my experience in planning methods for Deep Reinforcement Learning at Imperial College London as a guest speaker of Students for AI to inspire and influence their development in this field.

I was a co-founder and the first head of a student research club Gradient. I was managing the organisation, teaching Deep Learning and Reinforcement Learning and providing members with mentoring which result in members’ innovative projects and Gradient presence at big events like SAN Conference in Toruń or FOKA in Gdańsk.

At Quantum I’ve faced a problem of creating a face detector for a RaspberryPI platform. Using Deep Learning and TensorFlow I’ve managed to train the detector in Google cloud on WIDER FACES dataset and then deploy it using OpenCV in C++ backend of Quantum Sense technology. I've also worked on emotion classification from audio and video.

I’ve worked in Artificial Intelligence Product Group where I was responsible for development of a low level deep learning library for custom Intel hardware i.a. I’ve created a power virus program to test the performance of the chip under full load.

I’ve worked at Video and Parallel Computing Group where I was developing and maintaining the OpenCL user mode driver for Intel’s integrated GPU i.a. work groups scheduling component.