Plans for week 39

Dear all, welcome back to FYS-STK3155/4155.
Here's the weekly update with some messages below which could be of interest  to some of you.
Last week we discussed logistic regression and started with optimization methods. We will devote this week to several of these optimization methods and in particular on ways to estimate the gradients. This will lead us from the simple gradient descent to various variants of stochastic gradient descent.  We will also discuss how to make life less painful with algorithms like automatic differentiation. 

The material this week is covered by the slides for week 39 at https://compphysics.github.io/MachineLearning/doc/web/course.html. In addition, for a good discussion on gradient methods, we would like to recommend Goodfellow et al section 4.3-4.5 and chapter 8. We will come back to this in our discussion of Neural networks as well. See https://www.deeplearningbook.org/
For Stochastic Gradient Descent and implementations thereof  we recommend chapter 4 of Geron's text at https://www.oreilly.com/library/view/hands-on-machine-learning/9781492032632/

The plan for this week is thus as follows:
Lab: Wednesday (remember that from 815-12pm we are at F?397 on Wednesday, only for this week) and Thursday work on project 1. For those of you struggling with resampling techniques, please take a look at Hastie et al, chapter 7, sections 7.1-7.6 and 7.10-7.11 Hopefully this may be helpful. The text is at https://link.springer.com/book/10.1007/978-0-387-84858-7
Pay also attention to the fact scaling by subtracting the mean of your data for the Franke function is something you will see for Ridge and Lasso. Else, keep in mind that the results for say the bias-variance trade-off depende on the number of data points you have. Try to change the number of data points in your studies.

Thursday and Friday we will focus on optimization and gradient methods. Next week we start with neural networks and deep learning.

NOTE: Due to UngForsk we don't have access to our regular auditorium on Thursday. Thursday's lecture will thus be digital only via zoom. We are sorry for this.

Best wishes to you all,
Morten et al

///////////////////   Seminars and ML competitions, the latter with a prize of 1200 Euros!!

---   Competition (could be a variant for project 3

See https://www.nora.ai/competition/mapai-precision-in-building-segmentation/index.html
Text in Norwegian here
I ?r pr?ver vi oss p? en ny maskinl?ringskonkurranse - MapAI / KartAI, hvor m?let er ? segmentere bygninger i bilder tatt fra luften: https://www.nora.ai/competition/mapai-precision-in-building-segmentation/index.html. Vinneren vil f? en pengepremie p? 1200 euro og alle deltakerne vil bli invitert til ? submitte et metode-paper til Nordic Machine Intelligence. Vinneren vil tilslutt offentliggj?res p? Northern Lights Deep Learning conference i Troms?. 

I ?r som i fjor lurte jeg p? om denne maskinl?ringskonkurransen kanskje kunne v?re interessant for studentene dine og om kanskje dette kunne v?rt et alternativ til en av de 3 innleveringene studentene dine m? gj?re? I ?r tror jeg fristene passer bedre i forhold til semesteret for dine studenter da frist for algoritme-submission er 25 november og frist for submission av paper er 15 desember.

Hva tenker du om dette?

Mvh
Bj?rn-J  Contact him at b.j.singstad@fys.uio.no

====  Seminar of interest

Dear All, 
welcome to a new Computational Science seminar. This coming Friday we have the pleasure of having Daniel Schroeder (Simula and OsloMet) and Kaspara Skovli G?sv?r (former masterr of science student at the CS program, see her thesis at https://www.duo.uio.no/handle/10852/96708).

They will talk about

Title: Computational Science and the Spread of Harmful Conspiracy Theories in Online Social Networks
Abstract: Nowadays, since almost anyone can post on social media, a strict distinction between source and consumer is no longer evident. As a consequence we are exposed to an exponential growing flood of misleading information produced by an uncontrollable crowd of often clueless creators. Although it is well understood that the spread of misinformation leads to fatal consequences, it seems impossible to manually sort dangerous content from the sheer volume of data published on a daily basis. Natural language processing is widely used to automatically classify suspicious content. Here, the strategy is to create manually labeled training sets and train classifiers to detect the content of interest. However, even though these approaches significantly reduce the amount of manual labor required, machine learning models lack an understanding of context. Thus, features like humor or irony may not be taken into account, leading to miss-classifications. Due to these shortcomings, there is considerable motivation to explore other, more general detection methods. We aim for a more generic approach, exploiting not only the content but rather the underlying interactions within online social networks, to gain knowledge about the properties and dynamics of the spread of misinformation with harmful consequences on a societal scale. Specifically, we investigate the evolution of temporal networks induced by interactions between Twitter users during misinformation events.

Time: This coming Friday at 4pm, (September 30)
Place: Center for computing in Science Education, Department of Physics, 4th floor

Pizza and refreshments as usual and Daniel and Kaspara will also present possible thesis projects, plus possibly summerjobs and more.

You are all most welcome,
Best wishes for the week,
Morten