pilooch

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.

Let's wait and see how the new RTX cards with tensorcores do behave... Converting from and back to fp32 is welcome.

I do, with others, a lot of ML anomaly detection in the cyber security context. Deeply has interesting ideas, especially the encoded logs via lstm. The work was presented at a workshop at NIPS 2017.

One of the interesting facts we ve been able to measure empirically over the past few years is that the statistical anomalies' scores magnitude as reconstruction error are uncorrelated with the criticality of the anomaly in terms of security / threat.

This means that in practice SOC operators need to label on top of the anomaly detection and a supervised model can do the reranking after a while.

Figar, a simple enhancement to the deep net heads behind the RL mechanism fixes the jitter for most Atari games as far as I have tested it. See https://openreview.net/forum?id=B1GOWV5eg and https://github.com/jolibrain/manette for code and results (not the author of the paper).

Oops, missread for https://modelzoo.co/

I had the same reaction. I don't like it too much when sites copy up information and only link to original content at the bottom of the page.

The collection is good though, it's sad that it looks like it is stealing from the sources.

For deep learning, deepdetect can be useful in dev and prod phase.

We built our own 160 tflop (multi machine) GPU setup for ~20k€. Granted, it's not the same quality level as the hgx2, but that gives a reference point for computing power costs nowadays.

I run a machine learning company in France. Call it a startup or not, doesn't matter. Most of what I read in the comments here is not exactly true nor relies on statistics (sound like my story is...).

For companies acting within the science or innovative space, France is a haven. CIR is a tax return that covers 30% of R&D personnel, JEI is a status that halves the taxes on personnel for 8 years, and nullifies gov taxes the first year, halving them the second year. Add to this the region and gov backed projects that match 50% of your investment in a project.

The truth is it is much much cheaper than in the US, and probably unmatched in Western Europe. Now, all this preexisted Macron.

Some hurdles remain investment, low margins for companies around 20 employees and 2M yearly, they get stuck at that stage and lack of transparency.

One major pain is interacting with large French corporations, with payment delays unheard of in the US. This slows many startups down when it does not kill them. Once bootstrapped it seems to get better.

I'm not sure what a startup nation is not whether I'd like to live in it, but France is certainly in good position to fuel early small innovative & scientific companies, not more but not less :)

MDPs are the core holistic framework for decision under uncertainty, especially pomdps and dec-pomdps. That and the various forms of the Bellman equation are with your attention. RL is a model free MDP.

Dont do this in production without backup tools but we use unsupervised learning for anomaly detection. We're machine learnists so this is natural for us and we originally built the system for a mega Corp and made it open source (though not all parts are yet public). Catch is it's not real time but he system outputs outliers every few hours or daily depending of the size of your traffic (daily for around 50M proxy daily logs). Outliers are then reviewed by scripts and few operators / analysts.

Look at Unity ML agents if you haven't already.

Many thanks for the thorough answer! Will read and ask questions if any. We have customers we help quantize the models for their FPGAs and this may be an avenue we could test.

Interesting, any public pointer to your dissertation ?

You can setup a feature based image search engine or image object search engine with Open Source libraries, a few additional lines of Python, and a couple of pre-trained models, see https://github.com/beniz/deepdetect/tree/master/demo/imgsear... and https://github.com/beniz/deepdetect/tree/master/demo/objsear...

We use DeepDetect everywhere. I started coding it up in 2015 out of immediate need for my customers in production at the time and waited for something else to come out. But until now, we've sticked to it and customers ask for it when they see us using the pre-configured pipelines. Some run their production with the DD server. So three years later we're still putting down some time on improving it, and support for caffe2 is coming up among other things. It already has built in semantic search due to customer demand. Welcome any feedback an comments btw. Also we're hiring, contact us, we're at http://jolibrain.com/

No but you'll ready yourself for the (geometric) programming of the next twenty years or so.

C++ is the language of deep learning. Python is the scripting language on top of it.

Fact is though that instead of coding up an object detector, the machine is doing it now. This job is now being controlled by ML personnel, not (yet) your standard programmer. What is left is glue code and architecture, at least for some time.

Key to annotation is using pre trained or partially trained object detectors along with similarity search to quickly highlight unlabelled, similar objects in a dataset. Think of group labelling instead of labeling one object at a time. One of the open source implementations of object search out there is https://github.com/beniz/deepdetect/tree/master/demo/objsear... and that you can build your own group annotation tool around.

Fastext is faster but does not beat vdcnn and other deep architectures. It all depends on the use case really.

See https://github.com/jolibrain/dd_performances you can reach 1fps on image classification and objects detection tasks on rpie3 with custom caffe and models.

rcnn is two steps and ssd is single step.

See the Google blog post from last summer, https://research.googleblog.com/2017/08/transformer-novel-ne... A novel simplified architecture for sequences and translation.

After three attempts at designing an ML rest-lile API, I ended up writing the deepdetect's API. Never touched its fundamentals since mid 2015. Very very far from perfect but has served great til now!

ML is gonna be the next commodity on the stack. It takes time, but considering some of the complexity, I'm impressed by the speed at which this is becoming reality.

Let's minimize the loss

Research and experimentation with novel architectures has switched to pytorch everywhere I look. All colleagues and myself outside production pipelines have switched to pytorch. TF is good but too messy and hard for R&D, that's the price to pay for using a production-grade tool in research operations.

With all due respect you don't appear to know much about what you are commenting about. There is a huge community of research and applications around rnn and all other architectures. Pytorch has an extremely vivid and fast growing codebase, across all neural architectures and applications, it's remarkable actually. One reason might be because it is simple an effective, including easy debugging. Another is because research now is much focused on new possibly complex and exotic architectures, new optimizers, inner guts and behavior understanding​, including theoretically. And it appears pytorch makes that easy. Don't throw it away too fast as a DL enthusiast :)

Rasp3 running openelec / kodi, never looked back. All Open Source, not snooping on you :)