Home > Uncategorized > Weekly QuEST Discussion Topics and News, 3 Mar

Weekly QuEST Discussion Topics and News, 3 Mar

QuEST 3 March 2017

We need to continue our interactions on the idea of embodied cognition for example for issues we face in cyberspace (also there was a thread of discussion on the implications of the no free lunch theorem and its application to supervised learning and its implications for cyber defense that we might want to discuss).

.. paper on machine learning by David Wolpert extends directly to cyber defense interest. Theorem 2 on page 12 proves that “all heuristics that people have come up with for supervised learning … fail as often as they succeed” when they try to extend the knowledge obtained from a training set onto an unknown off-training-set future (examples on page 13).

 

..many efforts … violate Wolpert’s theorem by attempting to project a known past onto an unknown future. Not only are …these … approaches as likely to fail as to succeed, …

 

 

Different person on the thread: The “No Free Lunch” (NFL) Theorems of Wolpert and Macready have no place in any practical discussion of algorithms. The NFL Theorems state that, on average, no algorithm can perform equally well on all problems. As it is fair to say that every algorithm .. is tailored to the problem at hand, the NFL Theorems are irrelevant in the real world.

In their 1997 paper, Wolpert and Macready themselves describe their theorems:

“Intuitively, the NFL theorem illustrates that if knowledge of [the problem], perhaps specified through [the problem’s probability distribution], is not incorporated into [the algorithm], then there are no formal assurances that [the algorithm] will be effective. Rather, in this case effective optimization relies on a fortuitous matching between [the problem] and [the algorithm].

An excellent paper by Ho and Pepyne (“Simple Explanation of the No Free Lunch Theorem of Optimization,” 2001) states the same result this way:

“The No Free Lunch Theorem of Optimization (NFL T) is an impossibility theorem telling us that a general-purpose universal optimization strategy is impossible, and the only way one strategy can outperform another is if it is specialized to the structure of the specific problem under consideration.”

There are more detailed arguments, but I think the words of the authors themselves (Wolpert and Macready) refute any applicability of the NFL Theorems to any real-world algorithmic development, whether it is for machine learning, cyber defense, or any other problem domain.

 

Another person on the thread: ….  Basically, that statement says (independent of the Cyberspace application domain): If there is ZERO overlap between the training distribution and the test distribution than there should be little to no expectation that the ML algorithm should perform well (barring any side information).  In fact it should be expected to perform as well as random.

However, in practice, it is often the case that there is some level of overlap between data distributions or side-information available (i.e. both malware classes operate on an information system, made up of code, etc.).  The challenge in ML then is to develop appropriate models that are able to capture and exploit this overlap (however small) and generalize to different but overlapping domains (however small).  This challenge is one of the focus areas of ML.

I think it is true that providing provable guarantees on performance may not be possible in the general learning setting.  But that is OK, if the heuristics and developed models improve the state-of-the-practice.  However, I do think it is possible to provide guarantees as a function of the data distribution and overlap.  This type of work shows up in statistical machine learning relating to domain adaptation – where performance is bounded by the domain overlap.

Wolpert and Macready are very careful in defining the problem and applicability.  In fact follow-on papers they show where the NFL does not apply to ML cases:

1) http://www.cs.cmu.edu/~matts/Research/mstreeter_gecco_2003.pdf

2) https://ipvs.informatik.uni-stuttgart.de/mlr/papers/04-igel-NFL.pdf

3) https://pdfs.semanticscholar.org/7022/541216c2f6f894781bcd55a4cb545dbf81e9.pdf

 

 

Cap insertion into the thread – this is where if anyone has insights I should amend to my inserts I would be thrilled to hear them / discuss them:

Great discussion – attached sorry for the size is a chapter that documents the plenary talk I gave at the World congress on Computational Intelligence in 2001 – in it we prove a version of the NFL (no free lunch theorem) associated with intelligence amplification versus AI – the idea is in most of the interesting applications there are humans that consume/use the machine learning recommendation thus intelligence amplification versus replacing the human as in artificial intelligence –

The statements above are completely correct as usual … – we (my opinion) cannot abandon machine learning but have to use it with clear understanding of its limitations– using AI/ML is never a bad idea unless those using it do not understand the implications of how it was developed with respect to its capabilities and limitations

Just a reminder of the thread on embodied cognition and cyber:  The idea that instead of envisioning a central ‘brain’ for generating the understanding of the cyber environment a model along the lines of the Octopus might be more appropriate.  I only bring this back up to see if there are additional ideas that people came up with we need to consider.  We were also reminded of the discussions we’ve had on ‘split-brain’ people – people who have had their corpus callosum severed surgically so we want to bring that into the discussion.

This idea of embodied cognition is also relevant to our interest in organizations like ISIS using the web/social media to extend their impact.  A brief review of data analysis of that activity in a recent work will also be presented.

https://www.technologyreview.com/s/603626/data-mining-reveals-the-rise-of-isis-propaganda-on-twitter/?set=603629

  • Data Mining Reveals the Rise of ISIS Propaganda on Twitter
  • Twitter has closed 25,000 accounts that supported the terrorist organization ISIS.
  • An analysis of these tweets shows how ISIS emerged with a message of extreme violence.

  • One factor behind this rapid rise to power was ISIS’s use of social media, and Twitter in particular, to spread its ideas. And that raises some interesting questions: what have ISIS members and sympathizers been talking about on Twitter?  And why did that message prove so infectious?
  • Today we get an answer thanks to the work of Adam Badawy and Emilio Ferrara at the University of Southern California in Los Angeles.
  • These guys have analyzed some two million messages posted on Twitter by 25,000 members of ISIS. They say their analysis reveals important insights into the way the radical militant groups use social media and why its message spread so rapidly.

This is all part of our continuing discussion on representation – we’ve suggested this is the key to autonomy – to have agents with a representation or set of representations to facilitate deliberation / decision making for robustness and for a common framework when part of a human-machine joint cognitive solution.

Representation:  how the agent structures what it knows about the world – so for example its knowledge (what it uses to generate meaning of an observable) 

Reasoning:  how the agent can change its representation – the manipulation of the representation for example for the generation of meaning

Understanding:  application of relevant parts of the representation to complete a task – the meaning generated by an agent relevant to accomplishing a task in the estimation of an evaluating agent

Can a better representation be the missing link for autonomy – instead of the representation being generated by optimizing an objective function tied to MSE on classification – imagine the objective function tied to stability, consistency and usefulness –

Mean square error for classification as the objective function does not necessarily generate a representation that replicates the characteristics we consider critical for consciousness – we believe the constraints used by nature are

Stability

Consistency

Usefulness

We think we can achieve the first by using streaming video and enforcing the features extracted to not be bouncing / varying too much (assuming stabilized video) – unsupervised learning

We think we can achieve the second by doing video of similar locations and reward consistency – supervised learning

We think effectors / interactions with the environment – we have to close the loop around the representation to achieve the last – reinforcement learning

With respect to our interest in representation and objective functions we’ve asked the question recently and there has been some back channel conversation on if multiple agents use a related objective function and related machinery for acquisition of knowledge what can we say about their respective representations – so we want to have the opportunity to answer questions along these lines – the idea is to do fusion or multiple agent decision making will require some shared framework and we want to wrap our minds around the alternatives.

One of our main goals was to design a joint cognitive social media system focused on ‘mindfulness’ and thus a ‘context aware feed’ that provides some value –  how can QuEST agents facilitate the human getting into the ‘zone’ – the illusory apparent slowdown in time – we conjecture that the conscious perception of time is associated with the efficient ‘chunking’ of experiences

This topic came up this week in the QuEST news we posted – for example note specifically the last sentence in the excerpt below:

https://www.theregister.co.uk/2017/02/22/facebook_ai_fail/

Facebook scales back AI flagship after chatbots hit 70% f-AI-lure rate

‘The limitations of automation’

So it begins.

Facebook has scaled back its ambitions and refocused its application of “artificial intelligence” after its AI bots hit a 70 per cent failure rate. Facebook unveiled a bot API for its Messenger IM service at its developer conference last April. Facebook CEO Mark Zuckerberg had high hopes.

TenCent’s WeChat was the model. Although WeChat began life as an instant-messaging client, it rapidly evolved into a major platform for e-commerce and transactions in China. But it largely keeps any AI guesswork away from real users.  *** see second article below **

With Facebook’s bot API, Zuckerberg had joined a “chatbot arms race” with Microsoft CEO Satya Nadella. For Nadella, chatbots were “Conversations as a Platform,” or even the “third run-time” – as important to humanity as the operating system or the web browser.

Some experts fretted that if China opened up a lead in AI, the West would be doomed to lose World War 3. Others suggested that whichever superpower lost the AI arms race would relapse into a state of primitive technology feudalism.

However, as we reminded you recently, the reality of “artificial intelligence” is far from impressive, once it’s made to perform outside carefully stage managed and narrow demos. As stage one, Facebook’s AI would parse the conversation and insert relevant external links into Messenger conversations. So how has the experiment fared?

In tests, Silicon Valley blog The Information reports, the technology “could fulfil only about 30 per cent of requests without human agents.” And that wasn’t the only problem. “The bots built by outside developers had issues: the technology to understand human requests wasn’t developed enough. Usage was disappointing,” we’re told. Now it’s simply trying to make sense of the conversation. ** always the hole – meaning making **

There’s even a phrase you won’t have seen in many mainstream thinkpieces about AI, predicting a near future of clever algorithms taking middle-class jobs. Brace yourselves, dear readers. Facebook engineers will now focus on “training [Messenger] based on a narrower set of cases so users aren’t disappointed by the limitations of automation.”

Ah.

“Their discussions are much more grounded in reality now compared to last year,” said another person close to the Messenger developers. “The team in there now is finding ways to activate commercial intent inside Messenger. It’s much less about, ‘We’ll dominate the world with AI.'”

Analyst Richard Windsor describes Facebook as “the laggard in AI,” failing to match the results Google. “The problems that it has had with fake news, idiotic bots and Facebook M, all support my view that when Facebook tries to automate its systems, things always go wrong. The problem is not that Facebook does not have the right people but simply that it has not been working on artificial intelligence for nearly long enough,” he wrote recently.

In its exclusive, The Information also notes that Facebook has been grappling with what we call the “Clippy The Paperclip problem”: the user views the contribution by the agent, or bot, as intrusive.

———-  exactly where we are focused!

This reminds us of the prior discussions we’ve had on chunking – so we want to review the work of Cowan:

The magical number 4 in short-term
memory: A reconsideration
of mental storage capacity V3
Nelson Cowan

  • BEHAVIORAL AND BRAIN SCIENCES (2000) 24, 87–185
  • Abstract: Miller (1956) summarized evidence that people can remember about seven chunks in short-term memory (STM) tasks. However, that number was meant more as a rough estimate and a rhetorical device than as a real capacity limit. Others have since suggested that there is a more precise capacity limit, but that it is only three to five chunks. The present target article brings together a wide variety of data on capacity limits suggesting that the smaller capacity limit is real. Capacity limits will be useful in analyses of information processing only if the boundary conditions for observing them can be carefully described.

 

– thus a QuEST ‘wingman’ agent that helps the human formulate, recognize and exploit chunks would provide the insights to better respond to what may seem without it to be an overwhelming set of stimuli and evoke the zone illusion – thus our comment – a conscious recommender system facilitates the human decision maker getting into the zone

As part of the discussion on representation cap will present information from a recent article that we’ve not made it to yet:

Brain-Computer Interface-Based
Communication in the Completely Locked-In
State
Ujwal Chaudhary

Chaudhary U, Xia B, Silvoni S, Cohen LG,

Birbaumer N (2017) Brain±Computer Interface±

Based Communication in the Completely Locked-In

State. PLoS Biol 15(1): e1002593. doi:10.1371/

journal.pbio.1002593

  • Despite partial success, communication has remained impossible for persons suffering from complete motor paralysis but intact cognitive and emotional processing, a state called complete locked-in state (CLIS).
  • Based on a motor learning theoretical context and on the failure of neuroelectric brain-computer interface (BCI) communication attempts in CLIS, we here report BCI communication using functional near-infrared spectroscopy (fNIRS) and an implicit attentional processing procedure.
  • Four patients suffering from advanced amyotrophic lateral sclerosis (ALS)Ðtwo of them in permanent CLIS and two entering the CLIS without reliable means ofcommunicationÐlearned to answer personal questions with known answers andopen questions all requiring a “yes” or “no” thought using frontocentral oxygenation changes measured with fNIRS.
  • Three patients completed more than 46 sessions spread over several weeks, and one patient (patient W) completed 20 sessions.
  • Online fNIRS classification of personal questions with known answers and open questions using linear support vector machine (SVM) resulted in an above-chance-level correct response rate over 70%.
  • Electroencephalographic oscillations and electrooculographic signals did not exceed the chance-level threshold for correct communication despite occasional differences between the physiological signals representing a “yes” or “no” response. ** EEG not work **
  • However, electroencephalogram (EEG) changes in the theta-frequency band correlated with inferior communication performance, probably because of decreased vigilance and attention. If replicated with ALS patients in CLIS, these positive results could indicate the first step towards abolition of complete locked-in states, at least for ALS.

There was a related thread between some of us on ‘pain’ –

  • “Pain” – describes the unpleasant sensory and emotional experience associated with actual or potential tissue damage.
  • Includes – pricking, burning, aching, stinging and soreness
  • Assume serves important protective function
  • Some children born with insensitive to pain – severe injuries go unnoticed –
  • Some differences with other qualia – sense of urgency associated with it a sort of primitive quality associated with it – both affective and emotional components
  • Perception of pain influenced by many factors – identical stimuli in same agent can produce different ‘pain’ –
  • Anecdotes of wounded soldiers not feeling pain – until removed from battle field – injured athletes not experience the pain until after the competition
  • THERE IS NO ‘PAINFUL’ STIMULI – a stimuli that will produce the quale of pain in every agent independent of operating conditions
  • PAIN as in all qualia is NOT a direct expression of the sensory event – it is a product of elaborate processing

Some may have seen the news this week:

  • Forget the drugs, the answer to back pain may be Tai chi, massage

http://www.usatoday.com/story/news/nation-now/2017/02/14/forget-drugs-answer-back-pain-may-tai-chi-massage/97887446/

news-summary-43

Advertisements
Categories: Uncategorized
  1. No comments yet.
  1. No trackbacks yet.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: