TU Wien:Similarity Modeling 1 VU (Eidenberger)

Daten

Diese LVA wird nicht mehr von dieser Person angeboten, ist ausgelaufen, oder läuft aus und befindet sich daher nur noch zu historischen Zwecken im VoWi.

Vortragende	Horst Eidenberger
ECTS	3
Sprache	English
Links	tiss:188501

Zuordnungen

Masterstudium Logic and Computation
Masterstudium Visual Computing
Masterstudium Media and Human-Centered Computing
Masterstudium Medizinische Informatik

Inhalt

Machine learning basics with a focus on image and audio but presented in a sort of alternative way. The vocabulary and explanations used in this lecture are quite different from what is normally used in the ML-field. For example here is why the lecture is called Similarity Modelling: When you perform feature extraction/engineering you esentially build a model for your task (you select or transform some aspect of the data that is important while disregarding the rest) -> You then train a model with those extracted features -> When you then use the model to make predictions for a new input your model essentially compares how similar that new input is to the data it was trained on -> That's why it is called Similarty Modelling (even though you could also just call it plain old machine learning).

Ablauf

WT21: There are no regular lectures, just video lectures as block on Youtube (See youtube link under “Unterlagen”). You can watch the video at any time. Question hours was offered, but rarely used.

Student can choose between two examination modalities: Summarizing content from the videos or oral exam.

In the practical part, the videos of Muppet Show are presented and you should recognize certain characters like Kermit, Pigs, etc. using ‘classic’ Machine Learning and Deep Learning both via audio and images.

Benötigte/Empfehlenswerte Vorkenntnisse

WT21 Some knowledge of ML/DL is very helpful.

WT22/23 : You definetly need prior knowledge in machine learning and deep learning, while it technically is taught in the lecutre, the explanations there are not enough for someone starting from zero. It also helps to have some knowledge on how to work with image / audio data (e.g what a spectrogram is / how they are created)

Vortrag

There are no slides, the learning materials are videos where Prof. Eidenberger talks while drawing on a sheet of papers. Generally the lecture is quite interesting as it took a very different perspective to machine learning compared to other courses at the TU. I also liked the possibility to choose between an oral exam or summarizing the contents. I did the latter and got a good grade. Although I heard that using external images is not advised and leads to a worse grade.

Übungen

For both SIM 1 and SIM 2 the task is to find different characters of the mupped show with video and audio. In SIM 1 one can use easier methods than in SIM II. In both cases (ST 22), we had to apply a deep neural net and old-school feature extraction for both audio and video. Unfortunately, the experimental setup provided by Prof. Eidenberger was weird. There was no given train-val-test split, although he argued that submissions are graded with regard to performance. This is, of course, not scientific. Older submissions found on GitHub suggest that most people worked with data leakage, due to a random split, which makes no sense for video images. I would suggest not this, as data leakage does not seems to be taken into account for grading. Furthermore, our group got a grade deduction as we split the work, meaning that either one had a responsibility to cover, while still discussing problems we found. Although never stated in the exercise description, this reduced the grade by 1. All in all I found the grading highly unfair.

Prüfung, Benotung

noch offen

Dauer der Zeugnisausstellung

noch offen

Zeitaufwand

WT21: The Lab Course requires some work, but is all within the 3 ECTS

Unterlagen

WT21: Lectures and Lab Course: https://www.youtube.com/playlist?list=PLmphOgUcAYnS7o0ys4Tj-Q7iEXHUQpkd3

Tipps

WT21: It makes sense to do SIM1 and SIM2 both together, but you can also do SIM1 only.

Materialien

Diese Seite hat noch keine Anhänge, du kannst aber neue hinzufügen.