 You are not logged in | log in |
| You are not logged in | log in |
2021-12-16, 12:15, meet.google.com/ooi-zxye-dxa
Jan Ludziejewski (Uniwerystet Warszawski)
The OpenAI Jukebox was a groundbreaking model in sound generation and is still considered to be the state-of-the-art in the music modeling task. It consists of two separate networks, Vector Quantization Variational Autoencoder, which strongly compresses the raw waveform into a series of discret...
2021-12-02, 12:15, meet.google.com/ooi-zxye-dxa
Piotr Tempczyk (Uniwersytet Warszawski)
LIDL: Local Intrinsic Dimension estimation using approximate Likelihood
Understanding how neural networks work is one of the most important questions in machine learning research. Their performance is connected with the shape of the data manifold. The structure of this manifold can be explored with local intrinsic dimension (LID) estimat...
2021-11-18, 12:15, meet.google.com/ooi-zxye-dxa
Michał Zając (Uniwersytet Jagielloński)
Continual World: Continual learning meets reinforcement learning
First, I'll introduce the setup of continual learning. I'll talk about how various methods mitigate catastrophic forgetting and improve forward transfer. I'll also show what are the trade-offs and requirements for continual learning methods. In the second part, I will introduce our be...
2021-11-04, 12:15, meet.google.com/ooi-zxye-dxa
Konrad Czechowski (Uniwerystet Warszawski)
Subgoal Search For Complex Reasoning Tasks
I will present our publication accepted to NeurIPS 2021. We proposed a method that improves search guided by neural networks in combinatorially complex domains. Its key component is a learned subgoal generator that produces a diversity of subgoals that are both achievabl...
2021-10-21, 12:15, online seminar: meet.google.com/ooi-zxye-dxa
Łukasz Kuciński (IM PAN)
We will talk about our recent compositionality paper accepted at NeurIPS 2021. Communication is compositional if complex signals can be represented as a combi-nation of simpler subparts. In this paper, we theoretically show that inductive biases on both the training framework and the data...
2021-06-17, 12:15, google meet (meet.google.com/yew-oubf-ngi)
Sebastian Jaszczur
Sparsity in Efficient Transformers
Large Transformer models yield impressive results on many tasks, but are expensive to train, or even fine-tune, and so slow at decoding that their use and study becomes out of reach. We address this problem by leveraging sparsity. We study sparse variants for all layers in the Transformer and propos...
2021-05-06, 12:15, meet.google.com/yew-oubf-ngi
Piotr Kozakowski
Q-Value Weighted Regression: Reinforcement Learning with Limited Data
Sample efficiency is a major challenge in the current Reinforcement Learning (RL) systems. Another is robustness - it is hard to find one RL algorithm that will perform well in a variety of settings. I am going to present QWR - a novel RL algorithm that performs on-par with Soft Actor Critic (...
2021-03-25, 12:15, meet.google.com/yew-oubf-ngi
Bartłomiej Polaczyk
Loss landscape of deep neural networks
Finding the global minimum of a general non-convex function is in general an NP-complete problem. The optimization objective of deep neural networks (DNN) is not only non-convex but even non-smooth in case of ReLU activation function. Yet, practice suggests that even simple first-order methods such ...
2021-03-18, 12:15, meet.google.com/yew-oubf-ngi
Piotr Tempczyk
LIDL: Local Intrinsic Dimension estimation using Likelihood
We investigate the problem of local intrinsic dimension (LID)estimation. LID of the data is the minimal number of coordinates which are necessary to describe the data point and its neighborhood without significant information loss. Existing methods for LID estimation do not scale well...
2021-03-04, 12:15, meet.google.com/yew-oubf-ngi
Kajetan Janiak
Modelling rigid body dynamics with physics-informed neural networks
Functions describing physical systems of rigid bodies lie in a low-dimensional subspace of functions that neural networks can represent. If we could construct a NN architecture, that for any set of parameters yields dynamics of some rigid bodies’ system, then we could improve sample efficiency...
