Deep Learning Certificate Part II plans

I’ll definitely look into this!

1 Like

Maybe off topic, but I’d like to be better educated on choosing the right tool for the job. When is it appropriate to use deep learning? I’m worried about falling foul of the law of the instrument.

4 Likes

Very on topic @chris !

2 Likes

I would be great to have a review session for part 2. I don’t know about others, but I’ve fallen behind in 100% understanding all concepts but will work on understanding everything before Part 2!

8 Likes

Is it possible to model a ‘fake news’ detector as a kind of anomaly detection problem?

Also, it might be worthwhile in investigating whether we can leverage deep learning techniques to actually process and sanitize messy data?

2 Likes

If you can find us a nice labeled dataset, we could certainly do that. Although it’s nearly identical to the IMDB sentiment model we build in this course, so may not be that interesting…

I can’t wait to take the Part II course. However, I am not located in the San Francisco area. I am wondering whether it’s possible to join the Part II course remotely, without waiting till May when the MOOC version is available?

2 Likes

Yes, you can apply for the international fellowship. Send your CV and a letter explaining what you’re looking to get out of the course to info@(our domain name). It’ll be held Sunday nights at 6.30pm pacific, over Skype.

3 Likes

I agree with Ben’s suggestion, a brief revision especially for LSTM’s / sequential prediction would be great.

I would be great to have a review session for part 2. I don’t know about others, but I’ve fallen behind in 100% understanding all concepts but will work on understanding everything before Part 2!

The ‘revision’ class can also include a summary of the common pitfalls, and practical advice on fitting reproducible deep learning solutions. (e.g. like the advice Jeremy gives in one of the earlier lessons on preparing your experimental set-up using a sample set, being aware of when the model is not expressive enough (under-ftting), etc.).

Apart from that an example in predicting events or future levels for multi-dimensional time series (e.g. using a publicly available vital signs or ICU dataset) would also be very interesting.

1 Like

I’d be really interested in

  • Patterns matching for human behaviors. Input data could be clickstream/web traffic data or any kind of online or offline “user activity” data.
    The challenges are:

    1. Find person who’s behaviors is similar to the one I choose. (Discover most likely potential customers, or hidden malicious insider, or fraudsters)
    2. Find another group of people who’s behaviors are similar to the group I’ve chosen (discover hidden fraud rings, criminal groups who are trying to stay under the radar)
  • Adding rich textual annotations, descriptions and tags to images and videos (similar to what was shown here: https://www.usfca.edu/data-institute/certificates/deep-learning-part-one , scroll to 17:44) - making images and videos easily searchable.

Gleb

1 Like

i’d love to join this if it starts later this month. let me know if anyone is interested!

I would love to join too.

If you have any suggested datasets, that would be most helpful.

BTW, I’m not likely to do a review session in the course, since I’m hoping everybody will be doing their own study between the courses; but we will be building on everything we’ve learnt, and I will be reminded people of what the various pieces are as we come across them.

1 Like

Excited to know about the international fellowship option. Is the international fellowship priced the same as the in-classroom course? Is it going to cover the same material at the same pace as the in-classroom one?

1 Like

@jeremy

If you have any suggested datasets, that would be most helpful.

I think the most obvious and comprehensive is the MIMIC dataset. I realised there was actually a recent Kaggle competition using this dataset - but I think it was private entry.

The data set is theoretically openly available but in practice there are some hurdles to clear for access.

If that is not feasible, another option for a medium-sized and labelled multivariate time series data set has appeared in the Kaggle Sales forecasting competiton. Unfortunately, the context is no longer medicine, but I think any time series prediction insights you would give are directly applicable to medicine. It may be too commercial for the course though (?).

I’m not sure if this applies, but something along the lines of voice recognition. Being able to train a model to identify different voices/sounds in a recording and/or understand spoken commands.

1 Like

No - it’s free. It covers the same material at the same pace.

4 Likes

It just so happens I’ve been studying that dataset for the last couple of days, and will definitely be using it! Great minds…

2 Likes

There doesn’t seem to be any datasets large enough and of high enough quality to do anything useful in this space, AFAIK. :frowning:

A +1 for active learning. Active learning can help with small data sets by selecting the right examples to evaluate. I think it helps to look at data also: why is the image new?