matpalm - Latest Comments

Re: brain of mat kelcey

AntonTon — Thu, 31 May 2018 05:14:33 -0000

Hey. This is very cool project. Could you share your train set somewhere? Just interested to tinker some experiments with different nets on this but have no bees:)

Re: brain of mat kelcey

L Ting — Tue, 29 May 2018 09:06:05 -0000

Awesome!!! With the declining Bee population around the world; your effort will certainly help scientists and farmers to better support Bee's ecosystem ;-) Thank-you so much!

Re: brain of mat kelcey

Simon — Tue, 22 May 2018 04:03:52 -0000

Btw: Please give us an update about patterns you find in the data!

Re: brain of mat kelcey

Simon — Tue, 22 May 2018 03:59:25 -0000

Super cool thing!
Are you in contact with bee or insect researchers?
This might be a very useful tool for any kind of insect population study.

Re: brain of mat kelcey

EyesOnHives — Mon, 21 May 2018 19:17:19 -0000

Wonderful write-up Mat! Curious which systems you came across "after a little research i discovered it seems noone has a good non intrusive system for doing it yet"?

We certainly share the same set of interests :)

Re: brain of mat kelcey

mat kelcey — Sat, 19 May 2018 17:54:08 -0000

luckily it's never very cold here (on a global scale at least) :)

Re: brain of mat kelcey

kentbrew — Sat, 19 May 2018 11:21:01 -0000

This is neat! Do you have cold winters where you live? If so, you might want to consider installing something that won't let cold wind blow directly in through the side entrance. On my top-bar hives I always build the entrance at one end or the other; here is some video: https://www.youtube.com/wat...

Re: brain of mat kelcey

Steve Gulick — Fri, 18 May 2018 12:28:47 -0000

For compiling to the NCS - you might find the following links helpful. I haven't yet tried OpenVINO yet. If you try it I would be interested in your experience.
https://www.zdnet.com/artic...

https://software.intel.com/...

Re: brain of mat kelcey

mmparker — Thu, 17 May 2018 09:18:20 -0000

Nice! You might also be interested in this audio-based approach, which I heard about last summer: http://journals.plos.org/pl...

Re: simhash

Bill Conan — Tue, 16 Jan 2018 23:25:41 -0000

is using 32bit for the hash enough for documents?

I could imagine lots of hash conflicts, so each document should have a hash close to 11111111111111111111111111111111?

Re: simhash

Nathan Miller — Mon, 16 Oct 2017 16:45:55 -0000

How do you go from the hash values (i.e.
"th".hash = -502157718
"he".hash = -369049682
...
)
to the binary strings (i.e.
1 37586 1001001011010010
...
)
?

Re: brain of mat kelcey

Jem Mawson — Mon, 24 Jul 2017 22:02:38 -0000

Welcome back Mat. I wish you all the best on the farm. Sounds great.

Re: brain of mat kelcey

Alexandre Lacoste — Sat, 17 Jun 2017 10:04:31 -0000

That must be hard! Really happy for the family though :)
Remote work at Element AI would definitely be possible!

Re: brain of mat kelcey

Jeff Dalton — Fri, 16 Jun 2017 19:29:45 -0000

Best of luck. I hope we can catch up before you leave! Just sent you a lunch invitation.

Re: brain of mat kelcey

Guest — Fri, 16 Jun 2017 16:14:33 -0000

Are you leaving because you have to (as in visa problems,etc) or is it a deliberate choice? couldn't understand from the writing, would love to hear the decision behind it.

Re: brain of mat kelcey

Rajesh Thevar — Thu, 15 Jun 2017 23:14:29 -0000

Best wishes on your next journey. And the farming seems to be an interesting idea. Farewell.

Re: brain of mat kelcey

Matt Sandy — Thu, 27 Apr 2017 21:00:41 -0000

High frequency of phrases being used in conjunction with a political candidate to find potentially related news sources working together to push a similar narrative would be interesting.

Re: brain of mat kelcey

mat kelcey — Wed, 26 Apr 2017 23:20:05 -0000

are you thinking of something specific?

Re: brain of mat kelcey

Matt Sandy — Wed, 19 Apr 2017 22:57:36 -0000

Might be worth looking into ngrams too.

Re: simhash

Charles Van Noland — Wed, 04 Jan 2017 03:40:29 -0000

So basically simhashing is about effectively 'averaging' all of the hashes of all of the shingles together? This is reminiscent of neural networks but even more so of Sparse Distributed Representations and Hierarchical Temporal Memory stuffs. Seems like there'd be some added benefit to also examining not just two-character shingles, but two, three, four, five, etc.. Or somehow combining them hierarchically, like a simhash of a simhash?

Re: my list of cool machine learning books

Rahul Singh — Sat, 29 Oct 2016 15:43:23 -0000

Can you state them too ?

Re: brain of mat kelcey

u8y7541 — Sat, 30 Apr 2016 14:45:40 -0000

Yay, an XKCD reader.

Re: brain of mat kelcey

Dorje Haoxi — Wed, 06 Apr 2016 23:28:20 -0000

Yes, I agree that a more complex simulator is wanted.
So far, my initial idea was that to use remote control to drive the car (by man), and after that, use a DQN described in http://arxiv.org/abs/1312.5602 to train the car. But, it comes an issue ( I guess): the car won't run faster than the man drive it because the training data were base on the man control history).
My goal is to use DQN that can improve the performance better and better rather than just close to what the man can do best. So, I thought maybe simulation with DQN would be much better way to achieve that. Then, the problem here becomes: in simulator we don't have real-time camera view of the track, we only have 2-d shape of the track.
So, I am thinking maybe hack a racing game like Foza.

Oh, by the way, we'd like use real car on real track in the future, so, Nvidia TK1 will be our new gear to play with. ( I wanted TX1, but it remains out of stock..)

Re: brain of mat kelcey

mat kelcey — Wed, 06 Apr 2016 16:41:06 -0000

Awesome! I actually started with the pi camera but changed to just sonars so I could focus on a simpler problem first. The main result I was going to try to reproduce is the domain transfer described in http://arxiv.org/abs/1505.0... I think having the bot in a simple controlled environment, something like a square with a couple of colored poles, could be reproduced in a ROS. It would have to be a more complex simulator than STDR, I imagine Gazebo will be fine. I'd started some simple data collection, Gazebo renderings vs PiCamera captures, but haven't done anything with it yet.

Re: brain of mat kelcey

Dorje Haoxi — Tue, 05 Apr 2016 00:53:14 -0000

Hey Mate, it's great to know your work.
It's very similar to ours, the only difference is we use camera instead of radars.

Since we use camera to do this job, it is very hard to use simulator to apply DQN in training.
Do you have any suggestions?

Cheers,
Dorje