COSMOS - Issues and Enhancements

This software has the possibility to be very good and the only one that has caught my eye, however there are some improvements which are very much needed:

1. Tool tips and status bar at bottom
Put some tool tips on things so we get a hint of what is what, and information in status bar for what it’s doing. I had no idea it was scanning till I watched my cpu in task manager for a good while. Then a day later found that right click on tray icon also shows scanning.

2. It’s extremely slow at scanning and classification
All I’ve seen is CPU usage of at most 25%, doesn’t use all cores and doesn’t touch the GPU. Now if your using ML you should be offloading this on to the GPU, especially for all of us with cuda cores available. I left it going for 24 hours, it still hasn’t completed… I gave up…

3. Not enough room for categories at the top
You might want to make this a scrollable area or wrap to max 2 lines and then scrollable or something. I resized the window and the category I was going to click on was gone, no idea how to find it again.

4. Galaxy map
Render via GPU, if already rendering via GPU you might want to optimise. I fed it around 300k samples, scan didn’t complete but dots are flickering in and out of existence unless you zoom right in.

5. Sample classification in galaxy map
I can’t see the use of anything other than instrument setting for me (maybe another based on genre/type). However your ML trained dataset needs more training, got a lot of samples in groups which they do not belong in.

6. List views
Show more per a page, would take forever to get to the end with large sample resources

7. Menu
Needs a config menu. With options such as:

  • List view items per a page
  • Enable CPU and/or GPU
  • Throttle CPU / GPU use for those who actually get this thing to use more than 25% of CPU or more than 0% of their GPU
  • Font size

8. What are our privacy rights with this?
Are our sample sets being sent to your server for classification, are you retaining a list of all samples on our computers? Are you selling this data?

9. Scanning icon
I had no idea what that circle with nothing in it next to my sample folder name added, then a single line appeared through it, I was still what is that? No tool tip… Who knows. Then after leaving it go for 24 hours, I realised it was filling up, a click filling up indicator. After 24 hours it only got about 1/3rd of the way through the 300k samples of which only 281.5k were detected. A tool tip or change this icon to a small font number with % at the end

10. Waves logo moving in folder list on left
What is that? does that mean it’s doing something?

11. Middle mouse button drag, but not play sample also.
See adobe photoshop, blender etc. Keep the middle mouse button down and drag to drag the galaxy map, its virtually useless in high density galaxy clusters. And so the hunt for some clear space and to get it to drag begins.

1 Like

On point 8, they can’t that would violate very stringent privacy laws which will only lead to being exposed, getting sued, being publicly shamed and the company doing badly as a result.

We’d also havre to provide the computer with authorisation before it would allow that to happen.

Although, if they wanted to implement data gathering for the sole purpose of training the AI to be more efficient and better than I’m all for it as we’d all benefit. That kind of data would be anonymous as personal information is irrelevant in that kind of data collection.

Well it took nearly 5 days to scan my 300k samples, galaxy map takes a good cup of tea to generate.

As for data rights and privacy, that was the question… Computer authorisation wouldn’t be needed, there are processes running in the background WavesPluginServer which could quite easily scrape this data, and be recorded, very simple process to get user path etc and crawl files and phone home, nvidia does it (used to scan all files on desktop at one time and if you had too many files / folder depths on the desktop part of the nvidia software would crash, now why scan all files on your desktop… And many more do things they shouldn’t I’m sure)…

As for training the ML dataset, audio is typically generated as an image and some image process is done to it (forgot what it is called at the moment, it brings out more detail on the waveform kind of like spectral analysis) then the ML process this, but a correct classification is required for end comparison result to be of use in training, which might not be good to leave in the hands of the users as people might classify 1 sound in multiple different ways, thus corrupting the dataset’s integrity. There also other tricks to dataset manipulation like trimming the sound/sound image at start / end / both ends / reversing it, cutting bits out of it and keeping the classification of it which can greatly enhance the accuracy of the algo.

But they really need to get the processing on the GPU as its ridiculously slow.

Every process gets “sandboxed” in a modern OS. Meaning it can’t even look into other folders without a user’s permission. Which is why we always get asked to give permission to the Document’s folder, or Downloads folder, out whatever. It certainly can’t go digging around secure perts of the system

Even if it could this info is encrypted, so they couldn’t do anything with that information anyway.

AI is a bit different in the sense that it learns. So the more kick drums it identifies, for example, it gets a more clearer idea of what constitutes a kick drum and is more capable of identifying kick drums the more it identifies. So it’s a learning process.

This training process is done by the developers, over a couple of months, the mote info it has to work with the better the results. So naturally, if we could extend the training to beyond the developer and use a kind of a “cloud computing” approach, its very probably the algorithm can become more efficient. This might even make for faster analysis.

Yeah its possible that few users could maybe taint the results. But I’m not sure I’d know a user that would in this case as it would be of a disadvantage to them and Cosmos will not be as effective as they need it to be. So I think that the integrity of this data pretty much self governs. Especially when you consider the potential scope of thousands upon thousands of users.

RE: Sandboxing
Well as a Developer I’m quite sure I can rip out any file from any drive using visalstudio application in c++ and c#, and most certainly with node electron based applications. As I only use linux and windows, there are protected folders but they are system related. I could quite easily crawl any drive where you would store sample or the appdata folder or common files folders on windows. The user folder and documents folder paths are often given freely to applications for data storage, however these days now favouring the appdata folder, but my documents is easily crawled by anything (well everything I’ve ever made in c# or electron node).

RE: Encrypted
Never applied it to any of my drives, different OS protect different folders with groups and permissions etc system/admin/group name etc, I would have to test to find out more… But not this month… Got work to do…

RE: Training process
ML training (fake AI) takes months and the datasets should be huge and the datasets should be modified/altered/mangled in various ways to give extra training data from already existing data available.

RE: Cloud training
Nvidia have some great stuff on this, never used it but can be scaled up exponentially.

RE: Training and detection speed
This should be part of the machine learning algorithm, as it can be instructed to try multiple times on the same data set to get the best accuracy with speed. Not to mention, it should be done on the GPU, eg google “Tensorflow GPU”, nvidia’s deep learning networks also do both.

RE: Tainted results / classification of samples added to training set
This would be the case where someone classifies for example top loops which is a super set of hat loops, cymbal loops, clap loops, rides loops etc, which again could be multi tier. Then there would be the classification of effects, someone calling something a zap or buzz could be a drone or fx/sfx etc and thus tainting the results but would take many entries to fully taint it, but then again see what happened to ML trained chat bots which were public user trained and ended up being shut down due to being racist, using foul language or just plain out being insulting (has happened to MS twice now if memory serves me correctly).

Big +1 for 11. It’s hard to navigate now. Also would be great to have different function for RMB. Now there is 3 mouse buttons with the same action.