pilooch

Agreed. AutoML is useful, but modeling and ensuring correct hypotheses are taken into account comes first.

See https://github.com/jolibrain/deepdetect/pull/641 that uses FAISS as a backend alternative to annoy (annoy supported as well). Deletion can be implemented by removing entries from the listing db while the vector remains within the index.

Tests show that FAISS is bit better than annoy on retrieval of both small and million items indexes. It also includes ind x compression techniques that in our tests do fair very well, with very low loss on mid size 500k image indexes.

As an ML + open source developer for over 10y, MacOS support for deep learning is already long gone, Linux is the prime AI/ML OS. However, Apple and NVidia parting away is a good omen for GPU competition I believe. Whatever Apple hw comes up with, if usable outside MacOS software stack, it'd be an interesting alternative.

What would be a good source of tetrachromat images ? Would love to train a model and "see" what these women see !

vs C++-14 ? Indeed most DL is in fact C++. The Pytorch recent C++ API is a must. As professionals in this industry, my colleagues and I have switched to full C++. I'd be interested in advantages of Rust vs C++ instead of Python (which truely in terms of performances is C in the background).

Very interesting, thanks for pointing to this. Gitlab appears to be doing a fantastic job a opening governance and writing FAQ-like content such as this one. Truely boosting the core values and interest of libre / open source software and it's business models.

"The beauty of Arm is that it is open. You can you can add all sorts of interesting technologies that will be beneficial to supercomputing, such as being tightly integrated. By marrying an Arm CPU with a Tesla GPU...".

The essence of it. NVidia is so close that it makes development very painful. At the moment it's the required path for deep learning, but hopefully it'll change a bit.

It was a public service company, France Telecom, and employees were looking at their job as being civil servants in a way. Management decided to force many changes on employees in the hope they would resign. Many committed suicide or got sick, depressed etc...

Exactly the reason why my colleagues and myself do all deep learning in C++, performance and portability, from cloud to RPie. We've even modified caffe2 so we could build the training graph from pure C++. We know this is not the current doxa :) It's also all open sourced just in case others might need it...

We do everything with https://www.deepdetect.com/. Small team and targeted applications, this may not apply to all needs. The time from dev to prod is very low though.

Very cool! The docker install is proprietary as well or require an API key right ? Looks like a very finetuned tool, I'll recommend it around.

I guess webgl makes things faster with hw acceleration ? We re using this https://github.com/alx/parasol that is based on sigma.js, usually along with ML. Does webgl allow to get a higher number of nodes and edges running fine in the browser ? Ability to get forceatlas within the browser is a plus in practuve weve found.

Looks very seriously made and documented, congrats ! Was looking at it a bit closely the other days and put it onto my list of future tools. There's been another somewhat related library released by facebook recently, https://ai.facebook.com/blog/open-sourcing-pytorch-biggraph-...

Interesting project! You could run the neural networks directly on the pi. Expect 3 fps for object detection, try ncnn or deepdetect + ncnn for ease of use. That's without GPU. With a Jetson Nano expect up to 50 fps. You could detect more classes as well, as needed.

Search is often formulated as 'learning to rank'.

This is a bold statement lacking any serious scientific basis wrt advancing the state of sensory AI (pattern recognition) and robotics (actuation).

Universities and public research facilities are the existing democratic research institutions across the world. How can you defend not simply funding them and letting democracy handle it ?

One reason I see as an open source developer is low touch adoption and environment control. Over with build support on many Linux flavors (redhat from the past, archlinux from the ghetto :) ), use dockerized version instead. This might be my narrow corner of the world but users have driven me to dockerize much of what I do code, and my support hassle has lowered, a lot.

What is the use of 100fps vision models other than being the input to a controller (e.g. driving, flying etc) ? A raspberry can hold up to 3fps with standard open source frameworks, and this is enough for many applications, e.g. construction site surveillance etc... Not criticizing, rather a genuine interest to understand the edge ML vision market.

Cool! We all set this across a spreadout VPN so that our apps can keep track of the GPU availability https://github.com/jolibrain/gpustat_server It does spit JSON out with memory and computer usage.

Interesting conversation. I would add that papers by Lecun and others have been using character based convolutions on pure text since 2015 with great success. VDCNN is still a very good way to go for classification, and is much faster to train than RNN due to effective parallelization.

On a side note, sad to see these conversations about SOTA deep learning to be so adversarial... You're wrong / you're right kinda thing. It's an empirical science mostly at the moment, surf the gradient, be right and wrong at the same time !

Many thanks to the author for this very informative write-up. We are putting our open source deep learning codebase to run on raspberry with good performances for us, near 3fps for image object detection but we had little insights on how things truly worked internally to the pi. Again, very interesting, thanks.

Well, if it's Open Source / free software, as the ML research field is, then it is a meritocracy.

There are many recent works originating from the CFD community now. Too many to list here, you can PM me if you don't find them easily online for some reason.

There's a new trend of using deeplearning to replace and mix the solvers. Some ongoing research looks at tackling the compressible flows. Accelerating lattice boltzman methods with DL is also under study at various labs. See https://github.com/jolibrain/fluidnet_cxx for a reimplementation of fluidnet with aten/Pytorch tensors.

The literature on DL + cfd is growing steadily with some interesting papers at machine learning conferences. We are seeing the first set of applications in industry as well, very exciting !

Wow, Jean-Pierre Aubin mentioned on first comment on HN, that makes me so happy. I have had him as a teacher 20 years ago, an incredible teacher, always grounding the maths into their history, and its social production, with descriptions of Banach life and others along with their motivations.

Thanks to Aubin, I did my PhD in AI a long time ago and am still working in the field now!

If someone from fair is around, what does this mean for caffe2 ? Is the new torch c++ API a form a replacement for caffe?

Sounds much like conditions to a successful solving of a problem when modeled with machine learning. Have enough data, immediate (as opposed to delayed) feedback, and enough structure in the data. That neural networks in our brains might be learning an embedding or two :)

Interesting DB for feature storage and LSH is good choice I believe. I'm wondering why the tight link to pytorch C++ tensors (under refactoring actually), bit I haven't looked at the euclidendb code yet. Thanks for sharing !

Those interested can also find an open source integration of lmdb + annoy here: https://github.com/jolibrain/deepdetect/blob/master/src/sims...

This the underlying support for similarity search based on embeddings, including images and object similarity search, see https://github.com/jolibrain/deepdetect/tree/master/demo/obj...

This is running for apps such as a Shazam for art, faster annotation tooling and text similarity search.

Annoy only supports indexing once, while hnwlib supports incremental indexing, something I'm looking at.

For a programmatic view we use https://github.com/jolibrain/gpustat_server

What do you mean exactly ? I've been running a machine learning / reinforcement / optimization service company since 2015 and our customers range from the largest EU corps to small and mid size startups. If you mean how many customers use AI/DL/RL in production right now, it's a smaller percentage of our customers, around 25% I'd say, but growing, slowly. Most is about research roadmaps and testing this testing that, though these can be yezr long projects.

DL integration into or as a replacement to pieces of software stacks, plus long term model updates, measurements, testing and avoiding regressions (very hard, think of ensuring all past samples are similarly predicted as before after a model update, that's beyond simply improving the accuracy) is difficult and standard practice is building up slowly.