Ads are targeted with machine learning, the only possible long term defence is fighting back with machine learning. It's already been done and both sides will still improve over the next few years. But in the end, much like spam, online ads will go away, automatically.
On the tech side, using character based convolutional networks already frees the defense from writing endless rules, image ads can be nuked with another convnet, there's no technological lock.
I'm interested in knowing more about your chips. If smart sensors are one of the next steps for DL, as it seems from the high demand I get for embedded DL into cameras, etc... the existing offer by Nvidia (TK and TX) is not enough still. We have had conversations with FPGA DL startup, some others based on Nvidia chips, so I'd be interested in yours if you can share here. Otherwise PM is at deepdetect. Exciting times :)
Deepdetect for both dev and prod, with the minimum code in front of it. This setup is definitely not able to accommodate all modern ML needs, but the fast and secure model update from dev to prod is the easiest for us. Disclaimer: DD author so the bias is very high, apologies for this, maybe my comment will remain useful to some.
Yes, as a plugin, and on its way to make it to the trunk AFAIK. I've integrated it into deepdetect recently because even as a beta it works well and complements GPU based DNNs fairly naturally. Deep learning practitioners are already well equipped with GPUs so having XGBoost run on them as well is a good bonus!
I believe what general_ai means is that running models is not the issue here, it's about the FPS. NVidia has special GPUs on the TX1 and TK1 for this. Ability to run a model is about having enough memory for it. Ability to apply a model to a real time task is about having the compute, which for most tasks the Pi doesn't have. IIRC Pete Warden had ported some low level ops to the Pi GPU a few years ago, a difficult task.
This is why it is likely that what Google has in store is a form of inferenve-bound co-processor resembling their TPU.
Many people know what they are talking about on this thread, you just need to pay attention I believe. There's high demand for embedded deep learning at the moment, and I've already shipped several systems for a variety of tasks. At the moment none could live at required speed on the Pi.
Look https://deepdetect.com for an Open Source serve that covers most functionalities for building a personal deep learning as a service and API platform. It's not perfect but does a good job at images classification, object detection and character based text processing. There are AMIs available actually.
Chromebags, they've been on my back for over ten years now, and I still own and use the three I've bought over the years. The only drawback is that they never wear out, so you have no real excuse to change!!
It varies :) Mores especially it depends on your final platform and programming language. Android is fine, pure C++ without Python is more difficult though doable (eg deepdetect server has support for it), many other wrappers exist around TF.
UNK usually comes from the final sampling step: the distribution of words contains a special token UNK to represent all words with insufficient statistics in the training corpus.
I ve switched from air running Ubuntu to XPS 13 dev edition with Ubuntu. Not everything beautiful but a decent switch, great power, autonomy and screen.
FTR you can just build your own very easily out of any of the open source DL frameworks out there and their derivatives. E.g. digits and deepdetect will give you an API along with the pre-trained models.
Comparison is wrong between tfidf on words and CNN char. You should use char ngrams along with LR and this will beat all your classifiers with high probability. This is because your CNN char does not have enough data to draw all the useful chat ngrams. Doing it as preprocessing and passing it to LR is in practice always better on small datasets. You can go one step forward and add layers and test an MLP on your char ngrams.
It depends whether the rented ML is key to your product / startup. Alternative is running your own mini service with API based on open source software and public models or slightly modified models. Look at digits, deepdetect, bit fusion, etc..
For those interested in these matters, the books by Masanobu Fukuoka (1) are a must. They may challenge some of the HN crowd views on progress, technology and religion, as they did definitely challenge mine, but they carry significative scientific truths. His focus on fighting desertification was both genius and ahead of his time.