Numenta algorithm is not online learning. It does process and learn streaming data, but internally batches it into phases, a process not shown to happen in the brain.
In the HTM model (presumably the Numenta algorithm you're referring to), synaptic weights are updated with every new data point in discrete time steps (as opposed to continuous). In that sense, HTM is an online learning model. There was an experimental implementation of Temporal Memory (one component in HTM) that batched up some of those operations into phases, but that still happened in a single time step and that implementation has since been phased out (pardon the pun).
Why would you call AI just learning? Marketing? Self-aggrandizement?
AI is getting machines to solve problems they haven't been explicitly programmed to solve. As it is, we do not have AI. We have some bits and pieces of it. Best Mr algorithms so far only solve problems they have been explicitly trained and tweaked to solve.
Online learning has been attempted before, with very limited success. Making an online learning network stable is an open problem. These tend to quickly overfit the problem and get stuck.
> AI is getting machines to solve problems they haven't been explicitly programmed to solve.
That's one possible definition of AI, and not a terribly good one – just this morning, gmail solved my problem "I don't have John's phone#" without ever being explicitly programmed to "find John's phone#".
It seems people will always redefine AI to exclude whatever advances are made. Even passing the Turing test will just mean we've build an exceptionally good chatbot.
So here's my definition: AI is an algorithm that gets distracted from its original purpose to argue about the definition of AI on the Internet.
...and now back to categorizing these pictures. If I see one more Ostrich I'm going to segfault so hard.
The machine which turns itself off (the ultimate or sometimes useless machine) is an old gag made by Marvin Minsky and Claude Shannon.
[1] https://en.wikipedia.org/wiki/Useless_machine
>turning the traditional machine learning ‘black box’ into a ‘clear box’ neural network where new learnings can happen on the fly, in real time and at a fraction of today’s computational cost (no retraining over the whole dataset required).
I thought a black box meant that we aren't clear on why it makes the decisions it makes?
That's correct. You set up the shape of the neural net and you decide what aggregation function the neurons will use, but the process is largely opaque.
Yep, I checked out at: "...rise of deep learning since 2013, more or less when Google’s X Lab developed a machine learning algorithm able to autonomously browse YouTube to identify the videos that contained cats".
First, this started in 2012. Second, it wasn't Google - it was when Krizhevsky et al published their seminal work. Realistically, Google was slow to adopt to GPUs at the time, which I understand even contributed to Prof Ng's departure. It was Baidu who launched the first large scale deep-learning based image search, well ahead of Google.
Google has certainly caught up, but nobody can say they started it (and be taken seriously).
Those are great links! But the 2008 Hinton paper would not be considered deep learning, it is classic neural nets. It makes no mention of CNNs or GPUs, which is what really got this all going back in 2012 with ImageNet / Krizhevsky.
The ImageNet paper is from 2012, not 2010. That's when the computer vision community really went "wow". IIRC, almost every entry in ImageNet 2013 was using CNNs.
> it is classic neural nets. It makes no mention of CNNs or GPUs
Is using a GPU "essential" for something to be deep learning? I'd always thought that the important part was some sort of hierarchical representation learning.
GPUs certainly help, in that you don't want to wait all day while your code does that, but they're not necessary.
Good call on the 2012 not 2010 date. I missed that. GPU are not requirements of deep nn. Hinton's pseudo-bayesian + ReLU approach was the last piece of the deep neural net functionality. CNNs dated back to 1995-1998 with LeCun and Bengio. Although GPUs do accelerate deep NNs enough to be feasible on image data (thanks to Ng).
Yeah, it makes the unsubstantiated claim that since this process isn't how brains work, it isn't the key to teaching machines on the fly. But that ignores the whole field of online learning, which is making slow progress on just that..
You're right, assuming linearity is just and oversimplification, I think Tsvi Achler's video here will be useful to understand better what the article is about https://www.youtube.com/watch?v=9gTJorBeLi8
37 comments
[ 2.6 ms ] story [ 79.7 ms ] threadIn the HTM model (presumably the Numenta algorithm you're referring to), synaptic weights are updated with every new data point in discrete time steps (as opposed to continuous). In that sense, HTM is an online learning model. There was an experimental implementation of Temporal Memory (one component in HTM) that batched up some of those operations into phases, but that still happened in a single time step and that implementation has since been phased out (pardon the pun).
For some additional literature on the topic, see: - "Why Neurons Have Thousands of Synapses, a Theory of Sequence Memory in Neocortex", http://journal.frontiersin.org/article/10.3389/fncir.2016.00... - "Continuous Online Sequence Learning with an Unsupervised Neural Network Model", http://www.mitpressjournals.org/doi/abs/10.1162/NECO_a_00893... - "The HTM Spatial Pooler: a neocortical algorithm for online sparse distributed coding", http://www.biorxiv.org/content/early/2016/11/02/085035.abstr...
AI is getting machines to solve problems they haven't been explicitly programmed to solve. As it is, we do not have AI. We have some bits and pieces of it. Best Mr algorithms so far only solve problems they have been explicitly trained and tweaked to solve.
Online learning has been attempted before, with very limited success. Making an online learning network stable is an open problem. These tend to quickly overfit the problem and get stuck.
That's one possible definition of AI, and not a terribly good one – just this morning, gmail solved my problem "I don't have John's phone#" without ever being explicitly programmed to "find John's phone#".
It seems people will always redefine AI to exclude whatever advances are made. Even passing the Turing test will just mean we've build an exceptionally good chatbot.
So here's my definition: AI is an algorithm that gets distracted from its original purpose to argue about the definition of AI on the Internet.
...and now back to categorizing these pictures. If I see one more Ostrich I'm going to segfault so hard.
I thought a black box meant that we aren't clear on why it makes the decisions it makes?
First, this started in 2012. Second, it wasn't Google - it was when Krizhevsky et al published their seminal work. Realistically, Google was slow to adopt to GPUs at the time, which I understand even contributed to Prof Ng's departure. It was Baidu who launched the first large scale deep-learning based image search, well ahead of Google.
Google has certainly caught up, but nobody can say they started it (and be taken seriously).
http://www.cs.toronto.edu/~fritz/absps/tics.pdf
Here is a google tech talk from 2007:
https://www.youtube.com/watch?v=AyzOUbkUf3M
Companies didn't pick it up until more recently. GPU-ification happened in 2009 with Ng's group:
http://robotics.stanford.edu/~ang/papers/icml09-LargeScaleUn...
And yes, Krizhevsky et al (Hinton's lab) applied GPU deep learning to ImageNet in 2010:
https://papers.nips.cc/paper/4824-imagenet-classification-wi...
The ImageNet paper is from 2012, not 2010. That's when the computer vision community really went "wow". IIRC, almost every entry in ImageNet 2013 was using CNNs.
Is using a GPU "essential" for something to be deep learning? I'd always thought that the important part was some sort of hierarchical representation learning.
GPUs certainly help, in that you don't want to wait all day while your code does that, but they're not necessary.