'Revolution' ****
**** *****,*** ******** **** ********** ** "* ********** In ***** *********", *********:
**** ******** ********** **** ***** **** can ***** ** ******* **.* ******* accuracy ** ******* *****, ***** ************ ******* would ******** ** ******* 95 *******.
************* **** ***** ***** are, **** ***** ***** is *** **** *** the **.* ******* ** being ******** *** *** mentioned.
*** **.*% ****** **********.
Irony ******
*** ***** ** *** the ***** ** *** opening ******** ** *** deep ******** *****:
*** ***** ********* ******** is ***** ******* *** the ************ ****** ** sustained ** * ****** ofpreviously ********* *** **** and delivering too little in the past. [emphasis added]
*** ****** "***** *** cannot ******** *** **** are ********* ** ****** it" ** ********* *********** here.
Robots
**** ** ****** ***, 'thought *******' ******** ******* *********** ************ *** ****** *** coming *** **** *** "move ** ****** **** happen ********** *******". ** assured:
** **** ** ** an ******** ********* ***** less **** * **** from ***
** ********, ** *** both ******** ***** *** *****. We *** ** ** 'entirely ********* *****' **** robots **** ****. *************, for *******, ************** **** ****** * ****** social ***** ****.
Timing ******
** ** **** ** see that ********** *** ************ will ****** ********** ** different ****** **** ******. The ******* **** ********** did *** *** ***** ******.
*** ****** ****** ** predictions.
***** *** / ***** by ***** ****** ********. Bad ********* *** ****. Frustration *** ******** *****.
Deep ******** **********
**** *** * *** important ******* *** * healthy ********** ** **** learning ** *********:
- ******* ************: **** ******** **** ** **** in ******* ***** ************ applications. ***** **** **** really **** *** ***** ones **** ****** ** still *******. ** ** almost ******* **** *** performance **** **** ************* ****** applications. "**** ********" ** poised ** ** ******** in ********* *** ***** to **********.
- ***** *********** ********: ***** **** ******** fundamentally ****** **** ********, its *********** **** ** ************* impacted ** **** ** is *** ** ********** **. *** '**** world' ** ******* *** as ***** ************ *** deployed ** ********* *******, regions *** ***** (*.*., people **** *********, ****** differently), ** ** ******* that *********** ******** **** be *****. **** ** not ******* ** *** easily, ******* ** ******* the **** ******** ************ can ** **-******* ** handle ***** *********.
- ******* ****** ***********: ****** ******* ***** most ******** **** ******* from * *** *********, the ******** *** *********** of **** ******** ***** surveillance ******* ** ***** to **** ************* ***** on *** *********** ***** (which ** ** ***** supply *******), ********* ********* and **** ********* ** videos *** ******* ****** ** train **. **** ** likely ** ****** ** significant *********** ********* ****** vendors **** ** *** of **** *** '**** Learning' ** *** '***'.
- ******** ***** ******: ****** ***********,*** **** ******** ******** **** significant ******** ***** ******.
Early *******
**** *** *** *****, it ** ********* ** remind ********* **** **** is ***** ***** **** for **** ******** ** video ************.
* *** ** **** is ***** ***** ** still ********* *****. *** example, ** *** ****** ISC **** **** *****, one **** ******** ******** was ******* *** *** accuracy / *********** ** ***** system. ***** * ****** or ** ** ******** it, *** ***** ****** that *** **** ****** kept ** ******* **. It ****** ***, **** were ****** ******* ** a **** ** *** off ******. ***** ***** ****** will **** ********** **** one *** *** * lot ** **** ******** is **** *****.
Healthy **********
**** **** ********, ** will **** * ****** of ***** *** **** learning ** ****** (**** technically *** ******** ****) for ** ** **** a *** ********** ** video ************. ****** ** early, ********** ***** *** ill-will ******* **** ***** analytic ******** ** *** 2000s, ***** ********* ******** deep ********'* ********** *** the ********** ********* ****** have ** **.
Comments (11)
Undisclosed #1
How do you see intellectual property coming into play regarding the growth and adoption of 'deep learning' in surveillance/security applications??
i.e. can an algorithm me patented?
Create New Topic
Undisclosed #1
hmmmm...
no, actually I am in agreement - I don't think that detecting statistical rarity is patentable either. Any 'large-enough' data set can be analyzed/parsed to detect statistical rarity. It's just simple math.
What peaks the interest of customers (i think) is the ability of some 'thing' that can analyze these data sets fast enough to create a practical use case for buying that thing.
As John points out in the piece above, the marketers in our industry seem to be following the same path they took when VCA first emerged - dropping that buzz word into everything they produce.
Increased computational speed accompanied by lowering of traditional power output on a chip will definitely be able to achieve some really cool things.... but if deep learning is over-hyped (as VCA was) as a panacea rather than tying the new capabilities to actually solving existing security problems, imo deep learning has the potential to shoot itself in the foot - like VCA companies originally did before they tied the capabilities to specific security needs.
Create New Topic
Slava H
I think DNN advesarial attacks should be mentioned as well as one of concerns for security industry.
Its got enough attention to spin its own area of research in DNN community. http://www.cleverhans.io/ .
Create New Topic
Slava H
Yes this is a theoretical exercise, how important this problem would become for security industry has to be seen.
I would guess majority of industry players are using pre-trained existing models such as ImageNet and changing input and output layer for their needs. Would love to see anyone creating and training their own models.
As a rule of thumb: Models that are easy to optimize are easy to perturb. Linear models lack the capacity to resist adversarial perturbation; Models trained to model the input distribution are not resistant to adversarial examples. Ensembles are not resistant to adversarial examples.
General characteristic of model such as linearity of loss function and limited input features that's what is being used to optimise adversarial attack, therefore argument of security by obscurity not working for this case. (Here the paper on this issue: https://arxiv.org/pdf/1412.6572.pdf)
We have ourselves experienced such issues with pre-trained, well-known public model generating false human detections in a loading dock from some pipes on a floor.
To be fair, GANs (generative adversarial networks) is an active area of research due to being a nice training technique and not because of attack vector.
Create New Topic