Beskjeder

The final exam will take place at ITS (Kjeller) in the "Auditorium" room down in the basement.

Please  use the following guide in preparation for the exam:

1. Curriculum is defined by the  lecture slides. All the slides except for the slides which were not discussed in the class are in the curriculum.  In particular, the following slides were not discussed in  the class and therefore not in the curriculum:  
        (a) Relation network, Prototypical network,  MANN and LSTM meta-learner (in Learning concepts, slides 28-31)
        (b) Reptile  (in  Learning concepts, slide 33)
        (c) Zero shot learning (in Learning concepts, slides 35-40  )
        (d) Dialog Systems and Question answering   (in Text sequence processing slides 58-68)
2. Some slides contain detailed and complicated architecture  diagrams  copy-pasted from published a...

You can find the examination questions from 2018 here, and answers+ marking scheme here. 

Please note that some stuff are not in this year's curriculum.

You can find the mock examination from 2018 here  and the answers here.

Please note that this mock examination is longer than the proper examination. Further, note that some of the stuff might not be relevant to curriculum of this year.

Hello everyone,

We are in the final week of the course for this year. The last lecture will be on Friday the 15th and the topic is tracking of objects in videos.

We have the last two exercise classes next week at Blindern 1300-1400 on the 19th  Tuesday (Ole-Johan Dahls hus Seminarrom Pascal) and at Kjeller 13:15-15:00 on the 20th Wednesday (ITS GR 402).

If you want to discuss any TEK5040 related matter,  these classes will give you the opportunity for that.  

The compulsory exercise 3 is available from  here .  Unpack the file using tar xvfz oblig3.tar.gz. 

Deadline for submission is 24th November 23:59 Hrs.

You may look at the suggested solution here. Most of you had the implementation correct, or almost correct. Unfortunately some of you missed the correction of the assigment about negating the policy loss. You then ended up trying to minimize the return, and not surprisingly, did not get the results you wanted! A few people used numpy operations in calculating the losses, this does not unfortunately not work as TensorFlow is then not able to calculate the gradients with respect to the loss.

Optional assignment on word sequences is available here.

The exercise class at Blindern on Tuesdays is suspended until the next (last) compulsory assignment will be out.

You can find the package here.

This assignment is optional, but help will be provided as usual on Wednesday 13:15:15:00

There will be no exercise class at Blindern on the 22nd of October.  However, the regular exercise class at Kjeller on the 23rd of October will not be affected.

Due to requests from several students, the deadline for the mandatory assignment 2 is extended to 20th October 23:59 Hrs. We offer help for the assignment at both Blindern (13:00-14:00, Tuesday)  and Kjeller (13:15- 15:00, Wednesday).

There is an error in the specification of the policy loss in Section 2.3. The term that is specified in Section 2.3 for the policy loss is what we want to maximize. The policy loss should be the negative of this.

Webpage for our lecture plan has not been functioning properly since the last week. We have reported the problem, but the problem is still there. You can find the lecture slides for 11. October here.

In TensorFlow you rarely have to write for loops, in most cases there is a function that can perform the operation on the entire batch. Some functions you may find useful for this assignment are:

- tf.math.minimum

- tf.clip_by_value

- tf.math.reduce_mean

Furthermore a lot of basic operations are overloaded such that e.g. tf.constant([0.5, 2.0]) ** 2) --> [0.25, 2.0]

The missing requirements.txt file from Mandatory Assignment 2 can be found here.

Due to request from some students, we offer additional exercise help on Tuesdays 13:00-14:00 Hrs. 

Rom: Ole-Johan Dahls hus: Seminarrom Pascal

Note that this is in addition to the regular class on Wednesdays 13:15-15:00 at ITS.  The new arrangement starts from  the 8th Oct. with the compulsory exercise on reinforcement learning.

Mandatory assignment 2 is available here. It is due Thursday October 17th, 23.59h. Note that the assignment is quite verbose, so even though it seems very long, your tasks should be accomplishable in reasonable time. That said, you probably want to start early. 

Slides for lecture tomorrow on learning concepts can be found here.

We will delay the mandatory assignment (about a week). In the meantime make sure that you are up to date with the previous exercises. In preparation for the assignment, make sure that you understand the lecture notes from the previous lecture and perhaps also take a look at the references therein.

As our eminent teaching assistant noted, there is an error in the reduce_sum in the UniformAttention class.

Incorrect line:

context_vector = tf.reduce_sum(tf.expand_dims(attention_weights,axis=-1)*

feature_vectors, axis=-1)
 

Correct line :

context_vector = tf.reduce_sum(tf.expand_dims(attention_weights, axis=-1)* feature_vectors, axis=1)
 

We should sum over the 49 spatial dimensions (axis=1), not over the 256-dimensional feature dimension (axis=-1)

The assignment on image captioning is now out, look at "Timeplan/Schedule".

The lectures are webcasted at http://webcast-edge.uio.no/liveedge/its-301/playlist.m3u8

Note that is doesn't work in all browsers. Chrome should work.

The assignment should now be up on https://devilry.ifi.uio.no/ . Deadline is Thursday September 12th at 23.59. Please note that you may deliver as many times as you want, we will only look at the final delivery.

Please deliver a zipped file.