GHOST in new semester

According to the arrangements made at the organizational meeting, GHOST will operate in six groups in the current semester

1. AlphaGo Group – a group that aims to understand the AlphaGo algorithm developed by Google DeepMind, which a few years ago defeated best human player in the game of Go
2. VIZDOOM Group – a group whose goal is to extend the VIZDOOM library for reinforcement learning
3. Probabilistic Programming Group – a group working on understanding the Probabilistic Programming paradigm and its use for machine learning
4. DeepRacer Group – a group which builds a miniature of an autonomous car controlled by deep neural networks. The group, in cooperation with the RAI scientific club, will submit a grant application and (if grant will be awarded) RAI will build a miniature of the car, and GHOST will design algorithms analogous to those used in self-driving cars.
5. GHOSTDay Group – a special group organizing a conference related to machine learning at the Poznan University of Technology
6. IJCAI Group – (for now we have only one person – join us! 🙂 group starting in the competition about adversarial examples

GHOST: Episode IV – A New Hope

GHOST is now nearing its fourth semester, and it is time for a revolution! Starting next week, meetings will be more interactive – instead of the usual ‘seminar’ format (where each week the presenter prepared a presentation, and lead a discussion on a chosen topic) we will split up into project teams. Each team will work independently towards a specified goal, and discuss progress and potential difficulties during weekly meetings. Participation will now require more active engagement from each GHOST member, and everyone will chose a project, whose topic and complexity suits them most.

If you are interested in participating in GHOST in the upcoming semester, please come to the organizational meeting, which will take place on Tuesday, the 26th of February, at 11:45AM, in room 423 WE. You’ll learn about the planned projects and get to sign up for your chosen team. If you’ve never participated in GHOST, come meet us on Tuesday – we’re always happy to invite new members and answer any potential questions!

GHOST Report 2017-2018

We are proud to present a report on the activities of GHOST Machine Learning group in the years 2017-2018 prepared for His Magnificence Rector of PUT for evaluation purposes.

PLinML 2018

From Friday to Monday GHOST attended PLinML―a Polish View on Machine Learning. The conference took part in Warsaw and covered a variety of theoretical and applied topics, presented by speakers form both academia and industry. Our favorite talks included From Pixels to Human-level Performance in Multi-Player First Person Games presented by Wojciech Czarnecki from DeepMind, and Understanding the Early Phase of Training Deep Neural Networks by Stanisław Jastrzębski from the Jagiellonian University. Part of our group also attended an all-day tutorial on Conversational AI.

The conference also included a talk from none other than our own vice-president―Faustyna Krawiec, on Neural Models for Semantic Code Search―work done during her internship at Microsoft Research Cambridge. GHOST’s presence at the conference was further highlighted by two posters―Faustyna’s on the same topic, and Marcin Konieczny’s―about distributed clustering of big datasets, which he worked on over the summer at the University of Ljubljana. Meanwhile Dominik Huss took part in a quiz and won a copy of Christopher Bishop’s Pattern Recognition and Machine Learning!

Imbalanced Learning

Last week we said goodbye to Learning to Rank, by discussing Thurstone’s models, the Plackett-Luce model, ListNet and the LeToR test.

During today’s meeting we introduced the problem of Imbalanced Learning, and discussed some initial ideas on how to deal with imbalanced classes in Machine Learning. We will be concentrating on this topic for the next few weeks and learn about measures for imbalanced classification and algorithm-level approaches to the problem.

If you’re interested in exploring Imbalanced Learning with us, we’ll be meeting as usual on Thursdays, 4:50PM in 423WE.

GHOST Special Session – Edge AI

Last Monday a GHOST Special Session on Edge AI took place. This time we hosted Michał Giełda from AntMicro who gave an interesting presentation entitled “Edge AI – new capabilities, tools and platforms for a new wave of connected devices”. We learned about new hardware architectures supporting deep neural networks such as Jetson Xavier, Movidius Myriad or TPUs. Michał also emphasized the need for open architectures such as Risc-V and show us some interesting projects they are working on e.g. SkyWall and compression of deep nets.

Learning to Rank

We started off the new semester with an exciting new topic  – Learning to Rank! During our first meeting we reviewed basic Information Retrieval models, discussed search engine performance measures and started exploring how Machine Learning techniques can be applied in the construction of ranking models for Information Retrieval systems.

This semester GHOST meetings are scheduled for 4:50PM each Thursday, but next week is an exception – instead of Thursday (All Saints’ Day), we meet on Monday 29th October, 4:50PM, as usual in 423WE.

PP-RAI ‘2018

Next week Poznan University of Technology will host a conference of AI scientists from organizations and academic centers from all over Poland. The First Polish Meeting on Research in Artificial Intelligence (PP-RAI) is organized to promote closer cooperation between key Polish AI organizations such as Polish Association of Artificial Intelligence, Polish Neural Network Society and Polish Group of Learning Systems. The aim of the congress will be, inter alia, to create an AI map of Polish research centers as well as a discussion on new forms of organization and cooperation of AI research teams. Members of GHOST participate in the organization of this event and keep their fingers crossed for its success!

GHOST Special Session – Embed All The Things! Neural…

Yesterday, our third GHOST Special Session took place! This time we invited Marek Wydmuch, a data scientist from OLX who gave a very interesting presentation about neural embedding methods. His presentation covered many very interesting methods, starting from the basics of deep learning and ending on the most popular neural embedding method, namely, word2vec. Also, everyone received awesome gadgets from OLX 😉 Thank you very much!

We also invite you to the continuation of Marek’s presentation on 7th July. The presentation will cover more recent & advanced models like StarSpace or doc2vec.

Our adventure with decision trees has begun!

Yesterday we have started a new series of our meetings about decision trees. We have commenced with asking ourselves why other machine learning methods than deep neural networks are needed. After a long discussion, we have moved to the description of a basic decision tree algorithm (namely ID3).

For everyone who has missed our meeting and still would like to join us, we put our presentation here. If you have any question drop us an email or visit our Facebook fan page.