Archive

Archive for June, 2017

Weekly QuEST Discussion Topics and News, 23 June

QuEST 23 June 2017

 

We will start this week by responding to any issues people want to discuss reference to our previous meeting topic – Can machines be conscious?

We then want to discuss the idea of the ‘knowledge centric’ view of making machines conscious.  What we mean by that is we define knowledge as what is being used by a system to generate meaning.  The current limiting factor in machine generated knowledge is the resulting meaning the machines make is not rich enough for understanding – we define understanding to be meaning associated with expected successful accomplishment of a task.  If we want to expand the tasks that our machine agents can be expected to acceptably solve we have to expand the richness of the meaning they generate and thus we have to increase the complexity of the knowledge they create.  What are the stages of increasing knowledge complexity that will lead to autonomy?  We want to brainstorm a sequence of advances that would lead to system of systems that demonstrate peer, task and cognitive flexibility.

That leads to consideration of how that knowledge is represented and the topic below:

The paper by Achille / Soatto UCLA, arXiv:1706.01350v1 [cs.LG] 5 Jun 2017

On the emergence of invariance and disentangling in deep representations

Lots of interesting analysis in this article but what caught my eye was the discussion on properties of representations:

  • In many applications, the observed data x is high dimensional (e.g., images or video), while the task y is low-dimensional, e.g., a label or a coarsely quantized location. ** what if the task was a simulation – that was stable, consistent and useful – low dimensional?**
  • For this reason, instead of working directly with x, we want to use a representation z that captures all the information the data x contains about the task y, while also being simpler than the data itself.
  • Ideally, such a representation should be
  • (a) sufficient for the task y, i.e. I(y; z) = I(y; x), so that information about y is not lostamong all sufficient representations, it should be
  • (b) minimal, i.e. I(z; x) is minimized, so that it retains as little about x as possible, simplifying the role of the classifier; finally, it should be

(c) invariant to the effect of nuisances I(z; n) = 0, so that decisions based on the representation z will not overfit to spurious correlations between nuisances n and labels y present in the training dataset

  • Assuming such a representation exists, it would not be unique, since any bijective function preserves all these properties.
  • We can use this fact to our advantage and further aim to make the representation (d) maximally disentangled, i.e., TC(z) is minimal.
  • This simplifies the classifier rule, since no information is present in the complicated higher-order correlations between the components of z, a.k.a. “features.”
  • In short, an ideal representation of the data is a minimal sufficient invariant representation that is disentangled.
  • Inferring a representation that satisfies all these properties may seem daunting. However, in this section we show that we only need to enforce (a) sufficiency and (b) minimality, from which invariance and disentanglement follow naturally.
  • Between this and the next section, we will then show that sufficiency and minimality of the learned representation can be promoted easily through implicit or explicit regularization during the training process.

As we mature our view of how to work to these rich representation it brings up the discussion point of QuEST as a platform:

 

I would like to think through a QuEST solution that is a platform that uses existing front ends (application dependent by observation vendors) and existing big-data back ends like systems that follow the standard Crisp-DM approach like Amazon Web services … , and possibly a series of knowledge creation vendors  –

 

Independent of the representation used by a front end system that captures the observables and provides them to the QuEST agent – it becomes the quest agent’s job to take them and create two uses for them – the first is put them in the form to be used by the big-data solution (structure them so they can be used in the CRISP-DM process to find if there exists experiences stored – something close enough to them to provide the appropriate response) and the second form has to be consistent with our situated / simulation tenets – so they are provided to a ‘simulation’ system that attempts to ‘constrain’ the simulation that will generate the artificially conscious ‘imagined’ present that can complement the ‘big-data’ response – in fact the simulated data might be fed as ‘imagined observables’ into the back end – I would like to expand on this discussion

news summary (59)

Categories: Uncategorized

Weekly QuEST Discussion Topics and News, 16 June

QuEST 16 June 2017

We want to pick up where we left off last week – we spent time last week laying

out the QuEST model – how to build an agent that replicates much of the

representational characteristics we see in conscious critters in nature (the model

of course has an intuitive/subconscious aspect + the conscious aspect) – we can

take the next step by reviewing:

Can Machines Be Conscious? Yes—and a new Turing test might prove it

By Christof Koch and Giulio

IEEE spectrum June 2008 pg 55-59

• Pressed for a pithy definition, we might call it the ineffable and enigmatic

inner life of the mind. But that hardly captures the whirl of thought and

sensation that blossoms when you see a loved one after a long absence,

hear an exquisite violin solo, or relish an incredible meal.

• Some of the most brilliant minds in human history have pondered

consciousness, and after a few thousand years we still can’t say for sure if

it is an intangible phenomenon or maybe even a kind of substance

different from matter.

– We know it arises in the brain, but we don’t know how or where in

the brain. We don’t even know if it requires specialized brain cells (or

neurons) or some sort of special circuit arrangement of them.

• …

• Our work has given us a unique perspective on what is arguably the most

momentous issue in all of technology: whether consciousness will ever be

artificially created.

• It Will! …. there’s no reason why consciousness can’t be reproduced in a

machine—in theory, anyway.

We will go through the arguments in this article and another one:

Attention and consciousness: two distinct brain processes

• Christof Koch1 and Naotsugu Tsuchiya2

http://www.sciencedirect.com 1364-6613/$ – see front matter . Published by

Elsevier Ltd. doi:10.1016/j.tics.2006.10.012

• TRENDS in Cognitive Sciences Vol.11 No.1

Our discussions from last week on constructing QuEST agents:

 QuEST is an innovative analytical and software development approach to

improve human-machine team decision quality over a wide range of

stimuli (handling unexpected queries) by providing computer-based

decision aids that are engineered to provide both intuitive reasoning and

“conscious” deliberative thinking.

 QuEST provides a mathematical framework to understand what can be

known by a group of people and their computer-based decision aids

about situations to facilitate prediction of when more people (different

training) or computer aids are necessary to make a particular decision.

– these agents will have as part of their representation an instantiation of our

guiding tenets for qualia – our Theory of Consciousness – in the ‘conscious’ parts

of the representation – thus they will be ‘conscious’ in the sense they will comply

with the characteristics in the Theory of Consciousness – they will experience the

world by instantiating a representation that is compliant with those tenets as well

as an intuitive representation that will be an instantiation of current best

practices of ‘big-data’ {see for example deep learning} – it is our position that

nature does that –

We will revisit the concept of self:

Self – as we mature our discussion on autonomy – we have to address the

idea of ‘self’ – and ‘self-simulation’ – from our recent chapter on ‘QuEST for

cyber security’

4.2 What is consciousness?

Consciousness is a stable, consistent and useful ALL-SOURCE situated simulation that is structurally

coherent. [2, 4, 23, 27, 35, 44] This confabulated cohesive narrative complements the sensory data

based experiential representation, the subconscious. [22, 42] The space of stimuli resulting in

unexpected queries for such a representation complements the space of unexpected queries to the

experiential based representation that is the focus of the subconscious. (Figure 5) The vocabulary of the

conscious representation is made up of qualia. [6, 7, 8, 17] Qualia are the units of conscious cognition.

A quale is what is evoked in working memory and is being attended to by the agent as part of its

conscious deliberation. A quale can be experienced as a whole when attended to in working memory by

a QuEST agent. Qualia are experienced based on how they are related to and can interact with other

qualia. When the source of the stimulus that is being attended to is the agent itself the quale of ‘self’ is

evoked. A QuEST agent that has the ability to generate the quale of self can act as an evaluating agent

to itself as a performing agent with respect to some task under some range of stimuli. This is a major

key to autonomy. A QuEST agent that can generate the quale of self can determine when it should

continue functioning and give itself its own proxy versus stopping the response and seeking assistance

4.3 Theory of Consciousness

Ramachandran suggested there are laws associated with qualia (irrevocable, flexibility on the output,

buffering). [29] Since we use the generation of qualia as our defining characteristic of consciousness we

can use his work as a useful vector in devising our Theory of Consciousness. The QuEST theory of

consciousness also has three defining tenets to define the engineering characteristics for artificial

conscious representations. These tenets constrain the implementation of the qualia, working memory

vocabulary of the QuEST agents. [43,32] Tenet 1 states the representation has to be structurally

coherent. Tenet 1 acknowledges that there is minimal awareness acceptable to keep the conscious

representation stable, consistent, and useful. Tenet 2 states the artificially conscious representation is a

simulation that is cognitively decoupled. [18, 19] The fact that much of the contents of the conscious

representation is inferred versus measured through the sensors provides enormous cognitive flexibility

in the representation. Tenet 3 states the conscious representation is situated. [9,10] It projects all the

sensing modalities and internal deliberations of the agent into a common framework where

relationships provide the units of deliberations. [25,26,31,45,46] This is the source of the Edelman

imagined present, imagined past, and imagined future. [12]

4.4 Awareness vs Consciousness

There is a distinction between awareness and consciousness. Awareness is a measure of the mutual

information between reality and the internal representation of some performing agent as deemed by some

evaluating agent. Consciousness is the content of working memory that is being attended to by a QuEST

agent. Figure 8 provides examples of how a system can be aware but not conscious and vice versa. In the

blind sight example the patient has lost visual cortex in both hemispheres and so has no conscious visual

representation. [5] Such patients when asked what they see, say they see nothing and that the world is

black. Yet when they are asked to walk where objects have been placed in their path they often

successfully dodge those objects. Verbal asking is responded to based-on information that is consciously

available to the patients. These patients have awareness of the visual information but no visual

consciousness. Similarly, body identity integrity disorder (BIIDs) and alien hand syndrome (AHS) are

examples of issues that

illustrate low awareness

while the patient is

conscious of the

appendages. Paraphrasing

Albert Einstein “imagination

is more important than

knowledge,” we state

consciousness is often more

important than awareness.

There will always be

limitations to how much of

reality can be captured in the

internal representation of the

agent, but there are no limits

to imagination.

Autonomy requires

cognitive flexibility. Cognitive

flexibility requires, at least part of,

the internal representation be a

simulation (hypothetical). (Figure 9)

Situation awareness (SA) is defined

by Endsley to be the perception of

elements in the environment within a

volume of time and space, the

comprehension of their meaning, and

the projection of their status in the

near future. [13] The concept of SA

is intimately tied to the mutual

information between the internal

representation, reality, and

awareness. On the other hand,

situation consciousness (SC) is a

stable, consistent, and useful ALL-SOURCE situated simulation that is structurally coherent. This last

constraint of being structurally coherent requires the SC representation only achieve enough mutual

information with reality to maintain stability, consistency, and usefulness.

Figure 9. Einstein Quote

Figure 8 Venn Diagram Awareness vs Consiousness

Figure 10. QUEST Agents for Autonomy

or as Cognitive Exoskeleton

Figure 10 captures a desired end

state for our work. We envision

teams of agents (humans and

computers) that can align since

designed with similar architectures.

These solutions are called wingman

solutions. The goal is to generate a

theory of knowledge. Such a theory

would estimate the situation

complexity of the environment and

be able to predict a set of agents,

humans, and computers that have a

situation representation capacity that

matches.

The second topic – pursuing the thread that we need some means to generate the

‘imagined’ present/past/future – is associated with a relatively recent article on

video prediction.

DEEP MULTI-SCALE VIDEO PREDICTION BEYOND

MEAN SQUARE ERROR

Michael Mathieu1, 2, Camille Couprie2 & Yann LeCun1,

arXiv:1511.05440v6 [cs.LG] 26 Feb 2016

ABSTRACT

Learning to predict future images from a video sequence involves the

construction of an internal representation that models the image evolution

accurately, and therefore, to some degree, its content and dynamics. This is why

pixel-space video prediction may be viewed as a promising avenue for

unsupervised feature learning. In addition, while optical flow has been a very

studied problem in computer vision for a long time, future frame prediction is

rarely approached. Still, many vision applications could benefit from the

knowledge of the next frames of videos, that does not require the complexity of

tracking every pixel trajectory. In this work, we train a convolutional network to

generate future frames given an input sequence. To deal with the inherently

blurry predictions obtained from the standard Mean Squared Error (MSE) loss

function, we propose three different and complementary feature learning

strategies: a multi-scale architecture, an adversarial training method, and an

image gradient difference loss function. We compare our predictions to different

published results based on recurrent neural networks on the UCF101 dataset

 

news summary (58)

Categories: Uncategorized

Weekly QuEST Discussion Topics and News, 9 June

QuEST 9 June 2017

We want to start this week by returning to defining the advancements required in knowledge creation to achieve autonomy, specifically from the perspective of can we define a sequence of steps in advancing complexity of the knowledge being created required to achieve flexibility in peer / task / cognition.  We will have the discussion under the realization that we need a solution that scales, we need to improve every decision and be able to do so without re-engineering the autonomy for each application.  We need a knowledge creation platform!  What will that mean?

An autonomous system, AS, is one that creates the knowledge necessary to remain flexible in its relationships with humans and machines, tasks it undertakes, and how it completes those tasks in order to establish and maintain trust with the humans and machines within the organization the AS is situated in.  

….    

Self – as we mature our discussion on autonomy – we have to address the idea of ‘self’ – and ‘self-simulation’ – from our recent chapter on ‘QuEST for cyber security’

 

4.2 What is consciousness?

Consciousness is a stable, consistent and useful ALL-SOURCE situated simulation that is structurally coherent. [2, 4, 23, 27, 35, 44]  This confabulated cohesive narrative complements the sensory data based experiential representation, the subconscious. [22, 42]  The space of stimuli resulting in unexpected queries for such a representation complements the space of unexpected queries to the experiential based representation that is the focus of the subconscious. (Figure 5)  The vocabulary of the conscious representation is made up of qualia. [6, 7, 8, 17]  Qualia are the units of conscious cognition.  A quale is what is evoked in working memory and is being attended to by the agent as part of its conscious deliberation.  A quale can be experienced as a whole when attended to in working memory by a QuEST agent.  Qualia are experienced based on how they are related to and can interact with other qualia.  When the source of the stimulus that is being attended to is the agent itself the quale of ‘self’ is evoked.  A QuEST agent that has the ability to generate the quale of self can act as an evaluating agent to itself as a performing agent with respect to some task under some range of stimuli.  This is a major key to autonomy.  A QuEST agent that can generate the quale of self can determine when it should continue functioning and give itself its own proxy versus stopping the response and seeking assistance

 

4.3 Theory of Consciousness

Ramachandran suggested there are laws associated with qualia (irrevocable, flexibility on the output, buffering). [29]  Since we use the generation of qualia as our defining characteristic of consciousness we can use his work as a useful vector in devising our Theory of Consciousness.  The QuEST theory of consciousness also has three defining tenets to define the engineering characteristics for artificial conscious representations.  These tenets constrain the implementation of the qualia, working memory vocabulary of the QuEST agents. [43,32]  Tenet 1 states the representation has to be structurally coherent.  Tenet 1 acknowledges that there is minimal awareness acceptable to keep the conscious representation stable, consistent, and useful.  Tenet 2 states the artificially conscious representation is a simulation that is cognitively decoupled. [18, 19] The fact that much of the contents of the conscious representation is inferred versus measured through the sensors provides enormous cognitive flexibility in the representation.  Tenet 3 states the conscious representation is situated. [9,10] It projects all the sensing modalities and internal deliberations of the agent into a common framework where relationships provide the units of deliberations. [25,26,31,45,46]  This is the source of the Edelman imagined present, imagined past, and imagined future. [12]  

4.4 Awareness vs Consciousness

There is a distinction between awareness and consciousness.  Awareness is a measure of the mutual information between reality and the internal representation of some performing agent as deemed by some evaluating agent.  Consciousness is the content of working memory that is being attended to by a QuEST agent.  Figure 8 provides examples of how a system can be aware but not conscious and vice versa.  In the blind sight example the patient has lost visual cortex in both hemispheres and so has no conscious visual representation. [5] Such patients when asked what they see, say they see nothing and that the world is black.  Yet when they are asked to walk where objects have been placed in their path they often successfully dodge those objects.  Verbal asking is responded to based-on information that is consciously available to the patients.  These patients have awareness of the visual information but no visual consciousness.  Similarly, body identity integrity disorder (BIIDs) and alien hand syndrome (AHS) are examples of issues that illustrate low awareness while the patient is conscious of the appendages.  Paraphrasing Albert Einstein “imagination is more important than knowledge,” we state consciousness is often more important than awareness.  There will always be limitations to how much of reality can be captured in the internal representation of the agent, but there are no limits to imagination.

Autonomy requires cognitive flexibility.  Cognitive flexibility requires, at least part of, the internal representation be a simulation (hypothetical). (Figure 9)

Situation awareness (SA) is defined by Endsley to be the perception of elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future. [13]  The concept of SA is intimately tied to the mutual information between the internal representation, reality, and awareness.  On the other hand, situation consciousness (SC) is a stable, consistent, and useful ALL-SOURCE situated simulation that is structurally coherent.  This last constraint of being structurally coherent requires the SC representation only achieve enough mutual information with reality to maintain stability, consistency, and usefulness.  

Figure 10 captures a desired end state for our work.  We envision teams of agents (humans and computers) that can align since designed with similar architectures.  These solutions are called wingman solutions.  The goal is to generate a theory of knowledge.  Such a theory would estimate the situation complexity of the environment and be able to predict a set of agents, humans, and computers that have a situation representation capacity that matches.

 

 

The second topic – pursuing the thread that we need some means to generate the ‘imagined’ present/past/future – is associated with a relatively recent article on video prediction.  

DEEP MULTISCALE VIDEO PREDICTION BEYOND

MEAN SQUARE ERROR

Michael Mathieu1, 2, Camille Couprie2 & Yann LeCun1,

arXiv:1511.05440v6 [cs.LG] 26 Feb 2016

 

ABSTRACT

Learning to predict future images from a video sequence involves the construction of an internal representation that models the image evolution accurately, and therefore, to some degree, its content and dynamics. This is why pixel-space video prediction may be viewed as a promising avenue for unsupervised feature learning. In addition, while optical flow has been a very studied problem in computer vision for a long time, future frame prediction is rarely approached. Still, many vision applications could benefit from the knowledge of the next frames of videos, that does not require the complexity of tracking every pixel trajectory. In this work, we train a convolutional network to generate future frames given an input sequence. To deal with the inherently blurry predictions obtained from the standard Mean Squared Error (MSE) loss function, we propose three different and complementary feature learning strategies: a multi-scale architecture, an adversarial training method, and an image gradient difference loss function. We compare our predictions to different published results based on recurrent neural networks on the UCF101 dataset

news summary (57)

Categories: Uncategorized

Weekly QuEST Discussion Topics and News, 2 June

QuEST 2 June 2017

 

A thread going for the last couple of weeks that we need to get to is associated with epiphany learning

https://www.sciencedaily.com/releases/2017/04/170417154847.htm

 

http://www.pnas.org/content/114/18/4637.abstract

 

The topic was proposed by our colleague Prof Bert P – and then that was also supported by our recuperating colleague Robert P – from Robert:

This so-called ‘epiphany’ learning is more commonly known as insight problem solving and the original report on the phenomenon was Wallas in 1926 (he called it ‘illumination’). There are many papers in the literature on insight, and a well-known 1995 edited book is really great. …

 

What has attracted me to study insight is that it represents meaning making in a way that is tractable because the meaning making (insight or epiphany) occurs suddenly–exactly at the time the person get the insight, we know they have made meaning (i.e., the insight can be taken as a sign denoting a solution to a problem). Also, Bob E. and I have argued recently that insight is an intuitive cognition phenomenon (occurs suddenly from unconscious processing).

 

If anyone wants background to this paper, I have a lot of articles on insight I can send…

 

From the paper – Computational modeling of epiphany learning
Wei James Chena,1 and Ian Krajbicha

 

PNAS j May 2, 2017 j vol. 114 j no. 18 j 4637–4642

 

Abstract –

 

  • Models of reinforcement learning (RL) are prevalent in the decision-making literature, but not all behavior seems to conform to the gradual convergence that is a central feature of RL. In some cases learning seems to happen all at once. Limited prior research on these “epiphanies” has shown evidence of sudden changes in behavior, but it remains unclear how such epiphanies occur.
  • We propose a sequential-sampling model of epiphany learning (EL) and test it using an eye-tracking experiment. In the experiment, subjects repeatedly play a strategic game that has an optimal strategy.
  • Subjects can learn over time from feedback but are also allowed to commit to a strategy at any time, eliminating all other options and opportunities to learn.
  • We find that the EL model is consistent with the choices, eye movements, and pupillary responses of subjects who commit to the optimal strategy (correct epiphany) but not always of those who commit to a suboptimal strategy or who do not commit at all.
  • Our findings suggest that EL is driven by a latent evidence accumulation process that can be revealed with eye-tracking data.

 

 

In our FAQ we address learning in general:

 

 

  • What is learning?  What is deep learning?

 

Learning is the cognitive process used to adapt knowledge, understanding and skills through experience, sensing and thinking to be able to adapt to changes.  Depending upon the approach to cognition the agent is using (its choice of a representation ~ symbolic, connectionist, …), learning is the ability of the agent to encode a model using that representation (the rules in a symbolic agent via deduction or the way artificial neurons are connected and their weights for a connectionist approach using backpropagation – gradient descent).  Once the model has been encoded it can be used for inference.  Deep learning is a machine learning paradigm that uses multiple processing layers of simple processing units each loosely modeled after neuron brain cells in an attempt to generate abstractions from data. Deep learning has received a lot of attention in recent years due to its ability to process image and speech data, and is largely made possible by the processing capabilities of current computers with modest breakthroughs in learning approaches.  Deep learning is basically a very successful big data analysis approach.

 

Another thread has to do with the engineering of QuEST agents using a combination of DL for the sys1 calculations and cGANs for the generation of the qualia vocabulary – recall one application we were pursuing in this thread was the solution to the chatbot problem – there is a news article this week associated with this thread:

 

  • Ray Kurzweil is building a chatbot for Google

 

12

 

  • It’s based on a novel he wrote, and will be released later this year

 

by Ben Popper  May 27, 2016, 5:13pm EDT

 

 

Inventor Ray Kurzweil made his name as a pioneer in technology that helped machines understand human language, both written and spoken. These days he is probably best known as a prophet of The Singularity, one of the leading voices predicting that artificial intelligence will soon surpass its human creators — resulting in either our enslavement or immortality, depending on how things shake out. Back in 2012 he was hired at Google as a director of engineering to work on natural language recognition, and today we got another hint of what he is working on. In a video from a recent Singularity conference Kurzweil says he and his team at Google are building a chatbot, and that it will be released sometime later this year.

Kurzweil was answering questions from the audience, via telepresence robot naturally. He was asked when he thought people would be able to have meaningful conversations with artificial intelligence, one that might fool you into thinking you were conversing with a human being. “That’s very relevant to what I’m doing at Google,” Kurzweil said. “My team, among other things, is working on chatbots. We expect to release some chatbots you can talk to later this year.

 

 

One of the bots will be named Danielle, and according to Kurzweil, it will draw on dialog from a character named Danielle, who appears in a novel he wrote — a book titled, what else, Danielle. Kurzweil is a best selling author, but so far has only published non-fiction. He said that anyone will be able to create their own unique chatbot by feeding it a large sample of your writing, for example by letting it ingest your blog. This would allow the bot to adopt your “style, personality, and ideas.”

Another aspect of this thread is the question of whether the addition of cGANs could provide better meaning to DL systems – we propose to investigate this by attempting to demonstrate robustness to ‘adversarial examples’.  

Does anyone have access to the data necessary to reproduce the ‘adversarial examples’ – we’ve been pushing in QuEST that the current big need is a richer form of ‘meaning’ – the adversarial examples demonstrate the disparity of meaning to a DL solution versus a person – although it seems trivial I was wondering if we trained a cGAN with the images used to train a DL classifier that would be fooled by an adversarial example – but we take that adversarial example and provide it to the cGAN before giving it to the DL classifier if we could pull the DL result back to the correct side of the decision boundary?

 

  1. First train a DL system for a set of images – recall the Panda / Gibbon …
  2. Use that same set of data to train a cGAN to generate ‘imagined’ versions of those images – with the conditioning being on the original image for each episode versus just noise
  3. Train the DL system (possibly a second DL classifier) to take the cGAN images in and ‘correctly’ classify them
  4. Generate an adversarial example – provide to the original DL system – show incorrect meaning –
  5. Present that adversarial example to the cGAN – take the output of the cGAN and provide to the DL system trained on cGAN images to see if the processing the cGAN does on the adversarial example eliminates some/all of the errors in classification

 

The thought here is although the GANs in general do not produce ‘high-fidelity’ imagined data – they may provide the essence (‘gist’) that is enough to do classification – and such a representation could complement a system that recognizes the details

 

From our colleague Bernard suggest this is still an unsolved problem – may 2017 paper:

 

A very recent paper provides a summary of proposed methods for dealing with adversarial examples and some recommendations for future approaches. Also has links to some code on previous attempts to defeating adversarial examples, and authors plans to upload their code at some point.

 

Adversarial examples are not easily detected: Bypassing ten detection methods – Nicholas Carlini / David Wagner

https://arxiv.org/pdf/1705.07263.pdf

Abstract


Neural networks are known to be vulnerable to adversarial exam-
ples: inputs that are close to valid inputs but classified incorrectly.
We investigate the security of ten recent proposals that are de-
signed to detect adversarial examples. We show that all can be defeated, even when the adversary does not know the exact parameters of the detector. We conclude that adversarial examples are significantly harder to detect than previously appreciated, and we propose several guidelines for evaluating future proposed defenses.

news summary (56)

Categories: Uncategorized