sushirain

The bot can gather this restaurant specific information over several conversations with the restaurant. This wasn't possible before. This domain isn't too wide.

You can use a character RNN that reads the fragmented input char by char, and feed its output to your NN.

Is it possible to create an adversarial example without access to the weights of a model, and without being able to forward many images through it?

Can autonomous vehicles work without deep reinforcement learning? I thought that things like negotiating entry into a crossroad required DRL.

Thanks. I found that when they transitioned to Hardware2/AP2 they cut out most autopilot functionality, and since then have been gradually releasing updates that bring back that functionality.

I couldn't find a reference to this:"Tesla no longer sells cars with automatic driving capability." Can you point to one?

Very interesting. I wonder if they tried to predict part-of-speech tags.

The article says 5 years old, so it's probably not deep learning.

This is the relevant chart:

http://www.jcmit.com/mem2015.htm

DIMM seems to go down slower lately.

Would Apple?

My summary of the slides: to achieve "common sense", Unsupervised Learning is the the next step. Use all information in the data for predicting everything you can including the past. Use adversarial networks to do that.

If you don't have financial obligations, you could quit. Otherwise, maybe it's better to join an open source project on the side for a year. It takes years to go from noob to proficiency in coding, machine learning, cryptography. Some software positions are very selective.

Deep learning is already being monetized, e.g., at Google.

Vehicle detection: detect any vehicle, even from the side, even if its shape is rare, even at dark, etc. This could be solved with data.

Control: when to yield, without watching the face of the other driver, etc.

Edge cases: obeying a police officer, yielding to an ambulance, cooperating with other cars.

Weather conditions are a sensor / computer vision /control problem. It is far from solved.

The difference between 98% and 99.999% is very difficult to solve, and it's not going to happen in the next 5 years. LIDAR can't help, for example, with obeying a police officer gesture.

> Son said that he teamed up with a Saudi Arabian group to raise a $100 billion investment fund as part of a plan to move to the next stage.

Also see: "SoftBank and Saudi Arabia plan $100bn tech fund" https://www.ft.com/content/0370a5f4-9191-11e6-8df8-d3778b55a...

Why not use Medium instead of twitter?

"Marcin Wichary, the lead typographer and designer at Medium, is obsessed with all things typeface, including typewriters. So when he chanced upon Museu de la Tècnica in a small Spanish town, nothing could have prepared him for what he found." https://twitter.com/i/moments/792044988842598400

Can you be more specific?

The linguistic patterns that you mentioned (parts of speech, etc.) are too simplistic to describe language's idiosyncrasy and they ignore language acquisition.

Machine learning and data-driven methods require models with millions of parameters. Linguistic theories have traditionally not used such models.

Reminds me of Turing's morphogenesis. https://www.wired.com/2011/02/turing-patterns/

For example, ResNet from 2015 had 152 layers.

A real neuron takes in the order of 10ms to integrate and fire to the next neuron. Many subconscious reactions take less than 1 sec, which leaves time to a chain of length less than 100. Note that those neurons are not strictly arranged in layers.

The human visual cortex has 10^12 synapses [1]. One popular 2015 deep learning net (ResNet 152-layers) used 10^12 FLOPs to classify objects in one image (but less weights.)

In terms of depth, we're there. In terms of breadth, it will take several years. But the brain does things very differently. For example, it has top-down signals during "prediction."

[1] http://www.ncbi.nlm.nih.gov/pubmed/7244322

Your arguments are compelling. I hope others will embrace them more.

If we trained a net on many classes except for cats, then it would require a few examples of cat to recognize cats.

The difference from the 60's mis-predictions is that today we already have speech and visual object recognition systems working with 95% accuracy and we only need to get them to 99%. This gap will be closed by data collection and faster processors, no algorithmic leaps necessary.

Local minimum? In the next 10 years many deep learning applications will materialize: speech recognition will reach human level in production, autonomous trucks on highways, autonomous cars for consumers, AR, personal robot cleaners. The following 10 years will not fail us, too.

You do need deep-learning to get state-of-the-art OCR when the image is not scanned from a book, but photographed text (or text-in-the-wild) in unconstrained conditions.

It is a good example of the superiority of deep learning.

One recent example: Reading Text in the Wild with Convolutional Neural Networks. Jaderberg et al. 2016. https://www.robots.ox.ac.uk/~vgg/publications/2016/Jaderberg...

I don't think that AlphaGo was trained more on Lee Sedol's games, than on others' games. The team said that they can't find computer weaknesses until AlphaGo plays against top caliber.

Deep learning experts understand how neural nets work. But:

1. They don't usually try to explain the specific activations in the hidden layers of the network. This is hard and depends on the specific trained net.

2. They can't guarantee a net's performance before experimenting with it.

I think he meant recurrence when he wrote recursion, because recurrence is popular in deep learning (RNN) and recursion is not.