Archive for the ‘Machine learning’ Category

Things Speaking in Tongues

January 25, 2017


Speaking in tongues (aka Glossolalia) is the fluid vocalizing of speech-like syllables without any recognizable association with a known language. Such experience is best (not ?) understood as the actual speaking of a gutted language with grammatical ghosts inhabited by meaningless signals.

The man behind the tongue (Herbert List)

Do You Hear What I Say ? (Herbert List)

Usually set in religious context or circumstances, speaking in tongue looks like souls having their own private conversations. Yet, contrary to extraterrestrial languages, the phenomenon is not fictional and could therefore point to offbeat clues for natural language technology.

Computers & Language Technology

From its inception computers technology has been a matter of language, from machine code to domain specific. As a corollary, the need to be in speaking terms with machines (dumb or smart) has put a new light on interpreters (parsers in computer parlance) and open new perspectives for linguistic studies. In due return, computers have greatly improve the means to experiment and implement new approaches.

During the recent years advances in artificial intelligence (AI) have brought language technologies to a critical juncture between speech recognition and meaningful conversation, the former leaping ahead with deep learning and signal processing, the latter limping along with the semantics of domain specific languages.

Interestingly, that juncture neatly coincides with the one between the two intrinsic functions of natural languages: communication and representation.

Rules Engines & Neural Network

As exemplified by language technologies, one of the main development of deep learning has been to bring rules engines and neural networks under a common functional roof, turning the former unfathomable schemes into smart conceptual tutors for the latter.

In contrast to their long and successful track record in computer languages, rule-based approaches have fallen short in human conversations. And while these failings have hindered progress in the semantic dimension of natural language technologies, speech recognition have strode ahead on the back of neural networks fueled by increasing computing power. But the rift between processing and understanding natural languages is now being fastened through deep learning technologies. And with the leverage of rule engines harnessing neural networks, processing and understanding can be carried out within a single feedback loop.

From Communication to Cognition

From a functional point of view, natural languages can be likened to money, first as medium of exchange, then as unit of account, finally as store of value. Along that understanding natural languages would be used respectively for communication, information processing, and knowledge representation. And like the economics of money, these capabilities are to be associated to phased cognitive developments:

  • Communication: languages are used to trade transient signals; their processing depends on the temporal persistence of the perceived context and phenomena; associated behaviors are immediate (here-and-now).
  • Information: languages are also used to map context and phenomena to some mental representations; they can therefore be applied to scripted behaviors and even policies.
  • Knowledge: languages are used to map contexts, phenomena, and policies to categories and concepts to be stored as symbolic representations fully detached of original circumstances; these surrogates can the be used, assessed, and improved on their own.

As it happens, advances in technologies seem to follow these cognitive distinctions, with the internet of things (IoT) for data communications, neural networks for data mining and information processing, and the addition of rules engines for knowledge representation. Yet paces differ significantly: with regard to language processing (communication and information), deep learning is bringing the achievements of natural language technologies beyond 90% accuracy; but when language understanding has to take knowledge into account, performances still lag a third below: for computers knowledge to be properly scaled, it has to be confined within the semantics of specific domains.

Sound vs Speech

Humans listening to the Universe are confronted to a question that can be unfolded in two ways:

  • Is there someone speaking, and if it’s the case, what’s the language ?.
  • Is that a speech, and if it’s the case, who’s speaking ?.

In both case intentionality is at the nexus, but whereas the first approach has to tackle some existential questioning upfront, the second can put philosophy on the back-burner and focus on technological issues. Nonetheless, even the language first approach has been challenging, as illustrated by the difference in achievements between processing and understanding language technologies.

Recognizing a language has long been the job of parsers looking for the corresponding syntax structures, the hitch being that a parser has to know beforehand what it’s looking for. Parser’s parsers using meta-languages have been effective with programming languages but are quite useless with natural ones without some universal grammar rules to sort out babel’s conversations. But the “burden of proof” can now be reversed: compared to rules engines, neural networks with deep learning capabilities don’t have to start with any knowledge. As illustrated by Google’s Multilingual Neural Machine Translation System, such systems can now build multilingual proficiency from sufficiently large samples of conversations without prior specific grammatical knowledge.

To conclude, “Translation System” may even be self-effacing as it implies language-to-language mappings when in principle such systems can be fed with raw sounds and be able to parse the wheat of meanings from the chaff of noise. And, who knows, eventually be able to decrypt languages of tongues.

Further Reading

External Links

New Year: 2016 is the One to Learn

December 15, 2016

Sometimes the future is best seen through rear-view mirrors; given the advances of artificial intelligence (AI) in 2016, hindsight may help for the year to come.


Deep Mind Learning (J.Bosh)

Deep Learning & the Depths of Intelligence

Deep learning may not have been discovered in 2016 but Google’s AlphaGo has arguably brought a new dimension to artificial intelligence, something to be compared to unearthing the spherical Earth.

As should be expected for machines capabilities, artificial intelligence has for long been fettered by technological handcuffs; so much so that expert systems were initially confined to a flat earth of knowledge to be explored through cumbersome sets of explicit rules. But exponential increase in computing power has allowed neural networks to take a bottom-up perspective, mining for implicit knowledge hidden in large amount of raw data.

Like digging tunnels from both extremities, it took some time to bring together top-down and bottom-up schemes, namely explicit (rule-based) and implicit (neural network-based) knowledge processing. But now that it comes to fruition, the alignment of perspectives puts a new light on the cognitive and social dimensions of intelligence.

Intelligence as a Cognitive Capability

Assuming that intelligence is best defined as the ability to solve problems, the first criterion to consider is the type of input (aka knowledge) to be used:

  • Explicit: rational processing of symbolic representations of contexts, concerns, objectives, and policies.
  • Implicit: intuitive processing of factual (non symbolic) observations of objects and phenomena.

That distinction is broadly consistent with the one between humans, seen as the sole symbolic species with the ability to reason about explicit knowledge, and other animal species which, despite being limited to the processing of implicit knowledge, may be far better at it than humans. Along that understanding, it would be safe to assume that systems with enough computing power will sooner or later be able to better the best of animal species, in particular in the case of imperfect inputs.

Intelligence as a Social Capability

Alongside the type of inputs, the second criterion to be considered is obviously the type of output (aka solution). And since classifications are meant to be built on purpose, a typology of AI outcomes should focus on relationships between agents, humans or otherwise:

  • Self-contained: problem-solving situations without opponent.
  • Competitive: zero-sum conflictual activities involving one or more intelligent opponents.
  • Collaborative: non-zero-sum activities involving one or more intelligent agents.

That classification coincides with two basic divides regarding communication and social behaviors:

  1. To begin with, human behavior is critically different when interacting with living species (humans or animals) and machines (dumb or smart). In that case the primary factor governing intelligence is the presence, real or supposed, of beings with intentions.
  2. Then, and only then, communication may take different forms depending on languages. In that case the primary factor governing intelligence is the ability to share symbolic representations.

A taxonomy of intelligence with regard to cognitive (reason vs intuition) and social (symbolic vs non-symbolic) capabilities may help to clarify the role of AI and the importance of deep learning.

Between Intuition and Reason

Google’s AlphaGo astonishing performances have been rightly explained by a qualitative breakthrough in learning capabilities, itself enabled by the two quantitative factors of big data and computing power. But beyond that success, DeepMind (AlphaGo’s maker) may have pioneered a new approach to intelligence by harnessing both symbolic and non symbolic knowledge to the benefit of a renewed rationality.

Perhaps surprisingly, intelligence (a capability) and reason (a tool) may turn into uneasy bedfellows when the former is meant to include intuition while the latter is identified with logic. As it happens, merging intuitive and reasoned knowledge can be seen as the nexus of AlphaGo decisive breakthrough, as it replaces abrasive interfaces with smart full-duplex neural networks.

Intelligent devices can now process knowledge seamlessly back and forth, left and right: borne by DeepMind’s smooth cognitive cogwheels, learning from factual observations can suggest or reinforce the symbolic representation of emerging structures and behaviors, and in return symbolic representations can be used to guide big data mining.

From consumers behaviors to social networks to business marketing to supporting systems, the benefits of bridging the gap between observed phenomena and explicit causalities appear to be boundless.

Further Reading

External Links

Business Agility vs Systems Entropy

November 28, 2016


As already noted, the seamless integration of business processes and IT systems may bring new relevancy to the OOAD (Observation, Orientation, Decision, Action) loop, a real-time decision-making paradigm originally developed by Colonel John Boyd for USAF fighter jets.

Agility: Orientation (Lazlo Moholo-Nagy)

Agility & Orientation (Lazlo Moholo-Nagy)

Of particular interest for today’s business operational decision-making is the orientation step, i.e the actual positioning of actors and the associated cognitive representations; the point being to use AI deep learning capabilities to surmise opponents plans and misdirect their anticipations. That new dimension and its focus on information brings back cybernetics as a tool for enterprise governance.

In the Loop: OOAD & Information Processing

Whatever the topic (engineering, business, or architecture), the concept of agility cannot be understood without defining some supporting context. For OODA that would include: territories (markets) for observations (data); maps for orientation (analytics); business objectives for decisions; and supporting systems for action.

OODA loop and its actual (red) and symbolic (blue) contexts.

OODA loop and its actual (red) and symbolic (blue) contexts.

One step further, contexts may be readily matched with systems description:

  • Business contexts (territories) for observations.
  • Models of business objects (maps) for orientation.
  • Business logic (objectives) for decisions.
  • Business processes (supporting systems) for action.

The OODA loop and System Perspectives

That provides a unified description of the different aspects of business agility, from the OODA loop and operations to architectures and engineering.

Architectures & Business Agility

Once the contexts are identified, agility in the OODA loop will depend on architecture consistency, plasticity, and versatility.

Architecture consistency (left) is supposed to be achieved by systems engineering out of the OODA loop:

  • Technical architecture: alignment of actual systems and territories (red) so that actions and observations can be kept congruent.
  • Software architecture: alignment of symbolic maps and objectives (blue) so that orientation and decisions can be continuously adjusted.

Functional architecture (right) is to bridge the gap between technical and software architectures and provides for operational coupling.

Business Agility: systems architectures and business operations

Business Agility: systems architectures and business operations

Operational coupling depends on functional architecture and is carried on within the OODA loop. The challenge is to change tack on-the-fly with minimum frictions between actual and symbolic contexts, i.e:

  • Discrepancies between business objects (maps and orientation) and business contexts (territories and observation).
  • Departure between business logic (objectives and decisions) and business processes (systems and actions)

When positive, operational coupling associates business agility with its architecture counterpart, namely plasticity and versatility; when negative, it suffers from frictions, or what cybernetics calls entropy.

Systems & Entropy

Taking a leaf from thermodynamics, cybernetics defines entropy as a measure of the (supposedly negative) variation in the value of the information supporting the control of viable systems.

With regard to corporate governance and operational decision-making, entropy arises from faults between environments and symbolic surrogates, either for objects (misleading orientations from actual observations) or activities (unforeseen consequences of decisions when carried out as actions).

So long as architectures and operations were set along different time-frames (e.g strategic and tactical), cybernetics were of limited relevancy. But the seamless integration of data analytics, operational decision-making, and IT supporting systems puts a new light on the role of entropy, as illustrated by Boyd’s OODA and its orientation component.

Orientation & Agility

While much has been written about how data analytics and operational decision-making can be neatly and easily fitted in the OODA paradigm, a particular attention is to be paid to orientation.

As noted before, the concept of Orientation comes with a twofold meaning, actual and symbolic:

  • Actual: the positioning of an agent with regard to external (e.g spacial) coordinates, possibly qualified with the agent’s abilities to observe, move, or act.
  • Symbolic: the positioning of an agent with regard to his own internal (e.g beliefs or aims) references, possibly mixed with the known or presumed orientation of other agents, opponents or associates.

That dual understanding underlines the importance of symbolic representations in getting competitive edges, either directly through accurate and up-to-date orientation, or indirectly by inducing opponents’ disorientation.

Agility vs Entropy

Competition in networked digital markets is carried out at enterprise gates, which puts the OODA loop at the nexus of information flows. As a corollary, what is at stake is not limited to immediate business gains but extends to corporate knowledge and enterprise governance; translated into cybernetics parlance, a competitive edge would depend on enterprise ability to export entropy, that is to decrease confusion and disorder inside, and increase it outside.

Working on that assumption, one should first characterize the flows of information to be considered:

  • Territories and observations: identification of business objects and events, collection and analysis of associated data.
  • Maps and orientations: structured and consistent description of business domains.
  • Objectives and decisions: structured and consistent description of business activities and rules.
  • Systems and actions: business processes and capabilities of supporting systems.

Static assessment of technical and software architectures for respectively observation and decision

Then, a static assessment of information flows would start with the standing of technical and software architecture with regard to competition:

  • Technical architecture: how the alignment of operations and resources facilitate actions and observations.
  • Software architecture: how the combined descriptions of business objects and logic facilitate orientation and decision.

A dynamic assessment would be carried out within the OODA loop and deal with the role of functional architecture in support of operational coupling:

  • How the mapping of territories’ identities and features help observation and orientation.
  • How decision-making and the realization of business objectives are supported by processes’ designs.

Dynamic assessment of decision-making and the realization of business objectives’ as supported by processes’ designs.

Assuming a corporate cousin of  Maxwell’s demon with deep learning capabilities standing at the gates in its OODA loop, his job would be to analyze the flows and discover ways to decrease internal complexity (i.e enterprise representations) and increase external one (i.e competitors’ representations).

Further Readings

Brands, Bots, & Storytelling

May 2, 2016

As illustrated by the recent Mashable “pivot”, meaningful (i.e unbranded) contents appear to be the main casualty of new communication technologies. Hopefully (sic), bots may point to a more positive perspective, at least if their want for no no-nonsense gist is to be trusted.

(Latifa Echakhch)

Could bots repair gibberish ? (Latifa Echakhch)

The Mashable Pivot to “branded” Stories

Announcing Mashable recent pivot, Pete Cashmore (Mashable ‘s founder and CEO) was very candid about the motives:

“What our advertisers value most about
 Mashable is the same thing that our audience values: Our content. The
 world’s biggest brands come to us to tell stories of digital culture, 
innovation and technology in an optimistic and entertaining voice. As 
a result, branded content has become our fastest growing revenue 
stream over the past year. Content is now at the core of our ad 
offering and we plan to double down there.


Also revealing was the semantic shift in a single paragraph: from “stories”, to “stories told with an optimistic and entertaining voice”, and finally to “branded stories”; as if there was some continuity between Homer’s Iliad and Outbrain’s gibberish.

Spinning Yarns

From Lacan to Seinfeld, it has often been said that stories are what props up our world. But that was before Twitter, Facebook, YouTube and others ruled over the waves and screens. Nowadays, under the combined assaults of smart dummies and instant messaging, stories have been forced to spin advertising schemes, and scripts replaced  by subliminal cues entangled in webs of commercial hyperlinks. And yet, somewhat paradoxically, fictions may retrieve some traction (if not spirit) of their own, reprieved not so much by human cultural thirst as by smartphones’ hunger for fresh technological contraptions.

Apps: What You Show is What You Get

As far as users are concerned, apps often make phones too smart by half: with more than 100 billion of apps already downloaded, users face an embarrassment of riches compounded by the inherent limitations of packed visual interfaces. Enticed by constantly renewed flows of tokens with perfunctory guidelines, human handlers can hardly separate the wheat from the chaff and have to let their choices be driven by the hypothetical wisdom of the crowd. Whatever the outcomes (crowds may be right but often volatile), the selection process is both wasteful (choices are ephemera, many apps are abandoned after a single use, and most are sparely used), and hazardous (too many redundant dead-ends open doors to a wide array of fraudsters). That trend is rapidly facing the physical as well as business limits of a zero-sum playground: smarter phones appear to make for dumber users. One way out of the corner would be to encourage intelligent behaviors from both parties, humans as well as devices. And that’s something that bots could help to bring about.

Bots: What You Text Is What You Get

As software agents designed to help people find their ways online, bots can be differentiated from apps on two main aspects:

  • They reside in the cloud, not on personal devices, which means that updates don’t have to be downloaded on smartphones but can be deployed uniformly and consistently. As a consequence, and contrary to apps, the evolution of bots can be managed independently of users’ whims, fostering the development of stable and reliable communication grammars.
  • They rely on text messaging to communicate with users instead of graphical interfaces and visual symbols. Compared to icons, text put writing hands on driving wheels, leaving much less room for creative readings; given that bots are not to put up with mumbo jumbo, they will prompt users to mind their words as clearly and efficiently as possible.

Each aspect reinforces the other, making room for a non-zero playground: while the focus on well-formed expressions and unambiguous semantics is bots’ key characteristic, it could not be achieved without the benefits of stable and homogeneous distribution schemes. When both are combined they may reinstate written languages as the backbone of communication frameworks, even if it’s for the benefits of pidgin languages serving prosaic business needs.

A Literary Soup of Business Plots & Customers Narratives

Given their need for concise and unambiguous textual messages, the use of bots could bring back some literary considerations to a latent online wasteland. To be sure, those considerations are to be hard-headed, with scripts cut to the bone, plots driven by business happy ends, and narratives fitted to customers phantasms.

Nevertheless, good storytelling will always bring some selective edge to businesses competing for top tiers. So, and whatever the dearth of fictional depth, the spreading of bots scripts could make up some kind of primeval soup and stir the emergence of some literature untainted by its fouled nourishing earth.

Further Readings

Out of Mind Content Discovery

April 20, 2016

Content discovery and the game of Go can be used to illustrate the strengths and limits of artificial intelligence.

(Pavel Wolberg)

Now and Then: contents discovery across media and generations (Pavel Wolberg)

Game of Go: Closed Ground, Non Semantic Charts

The conclusive successes of Google’s AlphaGo against world’s best players are best understood when  related to the characteristics of the game of Go:

  • Contrary to real life competitions, games are set on closed and standalone playgrounds  detached from actual concerns. As a consequence players (human or artificial) can factor out emotions  from cognitive behaviors.
  • Contrary to games like Chess, Go’s playground is uniform and can be mapped without semantic distinctions for situations or moves. Whereas symbolic knowledge, explicit or otherwise, is still required for good performances, excellence can only be achieved through holistic assessments based on intuition and implicit knowledge.

Both characteristics fully play to the strengths of AI, in particular computing power (to explore playground and alternative strategies) and detachment (when decisions have to be taken).

Content Discovery: Open Grounds, Semantic Charts

Content discovery platforms like Outbrain or Taboola are meant to suggest further (commercial) bearings to online users. Compared to the game of Go, that mission clearly goes in the opposite direction:

  • Channels may be virtual but users are humans, with real emotions and concerns. And they are offered proxy grounds not so much to be explored than to be endlessly redefined and made more alluring.
  • Online strolls may be aimless and discoveries fortuitous, but if content discovery devices are to underwrite themselves, they must bring potential customers along monetized paths. Hence the hitch: artificial brains need some cues about what readers have in mind.

That makes content discovery a challenging task for artificial coaches as they have to usher wanderers with idiosyncratic but unknown motivations through boundless expanses of symbolic shopping fields.

What Would Eliza Say

When AI was still about human thinking Alan Turing thought of a test that could check the ability of a machine to exhibit intelligent behaviors. As it was then, available computing power was several orders of magnitude below today’s capacities, so the test was not about intelligence itself, but with the ability to conduct text-based dialogues equivalent to, or indistinguishable from, that of a human. That approach was famously illustrated by Eliza, a software able to beguile humans in conversations without any understanding of their meanings.

More than half a century later, here are some suggestions of leading content discovery engines:

  • After reading about the Ecuador quake or Syrian rebels one is supposed to be interested by 8 tips to keep our liver healthy, or 20 reasons of unsuccessful attempts at losing weight.
  • After reading about growing coffee in Ethiopia one is supposed to be interested by the mansions of world billionaires, or a Shepard pup surviving after being lost at sea for a month.

It’s safe to assume that both would have flunked the Turing Test.

Further Reading

External Links

AlphaGo & Non-Zero-Sum Contests

March 14, 2016

The recent and decisive wins of Google’s AlphaGo over the world best Go player have been marked as a milestone on the path to general artificial intelligence, one that would be endowed with the same sort of capabilities as its human model. Yet, such assessment may reflect a somewhat mechanical understanding of human intelligence.


Human intelligence goes well beyond winning zero-sum contests (Paolo Uccello)

What Machines Can Know

As previously noted, human intelligence relies on three categories of knowledge:

  1. Acquired through senses (views, sounds, smells, touches) or beliefs (as nurtured by our common “sense”). That is by nature prone to circumstances and prejudices.
  2. Built through reasoning, i.e the mental processing of symbolic representations. It is meant to be universal and open to analysis, but it offers no guarantee for congruence with actual reality.
  3. Attained through judgment bringing together perceptions, intuitions, and symbolic representations.

Given the exponential growth of their processing power, artificial contraptions are rapidly overtaking human beings on account of perceptions and reasoning capabilities. Moreover, as demonstrated by AlphaGo, they may, sooner rather than later, take the upper hand for judgments based on fixed sets (including empty ones) of symbolic representations. Would that means game over for humans ?

Maybe not, because human intelligence has evolved against survival stakes, not for games sake, and its innate purpose is to make fateful decisions when faced with unpredictable prospects: while machines look for pointless wins, humans aim for meaningful victories

What Animals Can Win

Left to their own, games are meant to be pointless: winning or losing is not to affect players in their otherwise worldly affairs. As a corollary, games intelligence can be disembodied, i.e detached from murky perceptions and freed from fuzzy down-to-earth rules. That’s not the case for real-life contests, especially the ones that drove the development of animal brains aeons ago; then, the constitutive and formative origins of intelligence were to rely on senses without sensors, reason without logic, and judgment without philosophy. The difference with gaming machines is therefore not so much about stakes as about the nature of built-in capabilities: animal intelligence has emerged from the need to focus on actual situations and immediate decision-making without the paraphernalia of science and technology. And since survival is by nature individual, the exercise of animal intelligence is intrinsically singular, otherwise (i.e were the outcomes been uniform) there could have been no selection. As far as animal intelligence is concerned opponents can only be enemies and winners are guaranteed to take all the spoils: no universal reason should be expected.

So, animal intelligence adds survival considerations to the artificial one, but it lacks symbolic and cooperative dimensions.

How Humans Can Win

Given its unique symbolic capability, the human species have been granted a decisive superiority in the evolution race. Using symbolic representations to broaden the stakes, take into account externalities, and build strategies for a wider range of possibilities, human intelligence clearly marks the evolutionary edge between human and other species. The combined capabilities to process non symbolic (aka implicit) knowledge and symbolic representations may therefore define the playground for human and artificial intelligence. But that will not take the cooperative dimension into account.

As it happens, the ability to process symbolic representations has a compound impact on human intelligence by bringing about a qualitative leap not only with regard to knowledge but, perhaps even more critically, with regard to cooperation. Taking leaves from R. Wright, and G. Lakoff, such breakthrough would not be about problem solving but about social destiny: what characterizes human intelligence would be an ability to assess collective aims and consequently to build non-zero-sum strategies bringing shared benefits.

Back to the general artificial intelligence project, the real challenge would be to generalize deep learning to non-zero-sum competition and its corollary, namely the combination and valuation of heterogeneous yet shared actual stakes.

However, as pointed by Lee Sedol, “when it comes to human beings, there is a psychological aspect that one has to also think about.” In other words, as noted above), human intelligence has a native and inherent emotional dimension which may be an asset (e.g as a source of creativity) as well as a liability (when it blurs some hazards).

Further Readings

External Links

AlphaGo: From Intuitive Learning to Holistic Knowledge

February 1, 2016

Brawn & Brain

Google’s AlphaGo recent success against Europe’s top player at the game of Go is widely recognized as a major breakthrough for Artificial Intelligence (AI), both because of the undertaking (Go is exponentially more complex than Chess) and time (it has occurred much sooner than expected). As it happened, the leap can be credited as much to brawn as to brain, the former with a massive increase in computing power, the latter with an innovative combination of established algorithms.


Brawny Contest around Aesthetic Game (Kunisada)

That breakthrough and the way it has been achieved may seem to draw opposite perspectives about the future of IA: either the current conceptual framework is the best option, with brawny machines becoming brainier and, sooner or later, will be able to leap over  the qualitative gap with their human makers; or it’s a quantitative delusion that could drive brawnier machines and helpless humans down into that very same hole.

Could AlphaGo and its DeepMind makers may point to a holistic bypass around that dilemma ?

Taxonomy of Sources

Taking a leaf from Spinoza, one could begin by considering the categories of knowledge with regard to sources:

  1. The first category is achieved through our senses (views, sounds, smells, touches) or beliefs (as nurtured by our common “sense”). This category is by nature prone to circumstances and prejudices.
  2. The second is built through reasoning, i.e the mental processing of symbolic representations. It is meant to be universal and open to analysis, but it offers no guarantee for congruence with actual reality.
  3. The third is attained through philosophy which is by essence meant to bring together perceptions, intuitions, and symbolic representations.

Whereas there can’t be much controversy about the first ones, the third category leaves room for a wide range of philosophical tenets, from religion to science, collective ideologies, or spiritual transcendence. With today’s knowledge spread across smart devices and driven by the wisdom of crowds, philosophy seems to look more at big data than at big brother.

Despite (or because of) its focus on the second category, AlphaGo and its architectural’s feat may still carry some lessons for the whole endeavor.

Taxonomy of Representations

As already noted, the effectiveness of IA’s supporting paradigms has been bolstered by the exponential increase in available data and the processing power to deal with it. Not surprisingly, those paradigms are associated with two basic forms of representations aligned with the source of knowledge, implicit for senses, and explicit for reasoning:

  • Designs based on symbolic representations allow for explicit information processing: data is “interpreted” into information which is then put to use as knowledge governing behaviors.
  • Designs based on neural networks are characterized by implicit information processing: data is “compiled” into neural connections whose weights (pondering knowledge ) are tuned iteratively on the basis of behavioral feedback.

Since that duality mirrors human cognitive capabilities, brainy machines built on those designs are meant to combine rationality with effectiveness:

  • Symbolic representations support the transparency of ends and the traceability of means, allowing for hierarchies of purposes, actual or social.
  • Neural networks, helped by their learning kernels operating directly on data, speed up the realization of concrete purposes based on the supporting knowledge implicitly embodied as weighted connections.

The potential of such approaches have been illustrated by internet-based language processing: pragmatic associations “observed” on billions of discourses are progressively complementing and even superseding syntactic and semantic rules in web-based parsers.

On that point too AlphaGo has focused ambitions since it only deals with non symbolic inputs, namely a collection of Go moves (about 30 million in total) from expert players. But that limit can be turned into a benefit as it brings homogeneity and transparency, and therefore a more effective combination of algorithms: brawny ones for actual moves and intuitive knowledge from the best players, brainy ones for putative moves, planning, and policies.

Teaching them how to work together is arguably a key factor of the breakthrough.

Taxonomy of Learning

As should be expected from intelligent machines, their impressive track record fully depends of their learning capabilities. Whereas those capabilities are typically applied separately to implicit (or non symbolic) and explicit (or symbolic) contents, bringing them under the control of the same cognitive engine, as humans brains routinely do, has long been recognized as a primary objective for IA.

Practically that has been achieved with neural networks by combining supervised and unsupervised learning: human experts help systems to sort the wheat from the chaff and then let them improve their expertise through millions of self-play.

Yet, the achievements of leading AI players have marked out the limits of these solutions, namely the qualitative gap between playing as the best human players and beating them. While the former outcome can be achieved through likelihood-based decision-making, the latter requires the development of original schemes, and that brings quantitative and qualitative obstacles:

  • Contrary to actual moves, possible ones have no limit, hence the exponential increase in search trees.
  • Original schemes are to be devised with regard to values and policies.

Overcoming both challenges with a single scheme may be seen as the critical achievement of DeepMind engineers.

Mastering the Breadth & Depth of Search Trees

Using neural networks for the evaluation of actual states as well as the sampling of policies comes with exponential increases in breath and depth of search trees. Whereas Monte Carlo Tree Search (MCTS) algorithms are meant to deal with the problem, limited capacity to scale up the processing power will nonetheless lead to shallow trees; until DeepMind engineers succeeded in unlocking the depth barrier by applying MCTS to layered value and policy networks.

AlphaGo seamless use of layered networks (aka Deep Convolutional Neural Networks) for intuitive learning, reinforcement, values, and policies was made possible by the homogeneity of Go’s playground and rules (no differentiated moves and search traps as in the game of Chess).

From Intuition to Knowledge

Humans are the only species that combines intuitive (implicit) and symbolic (explicit) knowledge, with the dual capacity to transform the former into the latter and in reverse to improve the former with the latter’s feedback.

Applied to machine learning that would require some continuity between supervised and unsupervised learning which would be achieved with neural networks being used for symbolic representations as well as for raw data:

  • From explicit to implicit: symbolic descriptions built for specific contexts and purposes would be engineered into neural networks to be tried and improved by running them on data from targeted environments.
  • From implicit to explicit: once designs tested and reinforced through millions of runs in relevant targets, it would be possible to re-engineer the results into improved symbolic descriptions.

Whereas unsupervised learning of deep symbolic knowledge remains beyond the reach of intelligent machines, significant results can be achieved for “flat” semantic playground, i.e if the same semantics can be used to evaluate states and policies across networks:

  1. Supervised learning of the intuitive part of the game as observed in millions of moves by human experts.
  2. Unsupervised reinforcement learning from games of self-play.
  3. Planning and decision-making using Monte Carlo Tree Search (MCTS) methods to build, assess, and refine its own strategies.

Such deep and seamless integration would not be possible without the holistic nature of the game of Go.

Aesthetics Assessment & Holistic Knowledge

The specificity of the game of Go is twofold, complexity on the quantitative side, simplicity on  the qualitative side, the former being the price of the latter.

As compared to Chess, Go’s actual positions and prospective moves can only be assessed on the whole of the board, using a criterion that is best defined as aesthetic as it cannot be reduced to any metrics or handcrafted expert rules. Players will not make moves after a detailed analysis of local positions and assessment of alternative scenarii, but will follow their intuitive perception of the board.

As a consequence, the behavior of AlphaGo can be neatly and fully bound with the second level of knowledge defined above:

  • As a game player it can be detached from actual reality concerns.
  • As a Go player it doesn’t have to tackle any semantic complexity.

Given a fitted harness of adequate computing power, the primary challenge of DeepMind engineers is to teach AlphaGo to transform its aesthetic intuitions into holistic knowledge without having to define their substance.

Further Readings

External Links