Human in the loop: a design pattern for managing teams working with ML

Human in the loop: a design pattern for managing teams working with ML
Human in the loop: a design pattern for managing teams working with ML Paco Nathan @pacoid R&D Group @ O’Reilly Media Strata CA San Jose, 2018-‐03-‐08
The reality of data rates “If you only have 10 examples of something, it’s going  to be hard to make deep learning work. If you have  100,000 things you care about, records or whatever,  that’s the kind of scale where you should really start  thinking about these kinds of techniques.” Jeff Dean Google  VB Summit (2017-‐10-‐23) venturebeat.com/2017/10/23/google-‐brain-‐chief-‐says-‐100000-‐ examples-‐is-‐enough-‐data-‐for-‐deep-‐learning/ 2
The reality of data rates Transfer learning aside, most DL use cases require   large, carefully labeled data sets, while RL requires   much more data than that. Active learning can yield good results with substantially smaller data rates, while leveraging an organization’s expertise to bootstrap toward larger labeled data sets, e.g., as preparation for deep learning, etc. reinforcement learning supervised learning active learning deep learning data rates (log scale) 3
The reality of data rates Transfer learning aside, most DL use cases require large much more Active learning smaller data rates, while leveraging an organization expertise to bootstrap toward larger labeled data sets, e.g., as preparation for deep learning, etc. reinforcement learning supervised learning active learning deep learning data rates (log scale) reinforcement learning supervised learning active learning deep learning data rates (log scale) active learning: indicated for many enterprise use cases 4
Why are AI programs different? 5 AI in the software engineering workflow  Peter Norvig Google  TheAIConf (2017-‐06-‐28) ▪ Content: models not programs ▪ Process: training not debugging ▪ Release: retraining not patching ▪ Uncertainty: of objective ▪ Uncertainty: of action/recommendation ▪ Uncertainty: propagates through model
Active Learning: case studies and patterns
Machine learning supervised ML: ▪ take a dataset where each element has a label ▪ train models on a portion of the data to predict the labels, then   evaluate on the holdout ▪ deep learning is a popular example,   but only if you have lots of labeled training data available 7
Machine learning unsupervised ML: ▪ run lots of unlabeled data through an algorithm to detect “structure” or embedding ▪ for example, clustering algorithms such as K-‐means ▪ unsupervised approaches for AI   are an open research question 8
Active learning special case of semi-‐supervised ML: ▪ send difficult decisions/edge cases   to experts; let algorithms handle routine decisions (automation) ▪ works well in use cases which have lots of inexpensive, unlabeled data ▪ e.g., abundance of content to be classified, where cost of labeling   is a major expense 9
Who’s doing this? 10
Design pattern: Active learning Real-‐World Active Learning: Applications and Strategies for Human-‐in-‐the-‐Loop ML  Ted Cuzzillo  O’Reilly Media (2015-‐02-‐05) Active learning and transfer learning  Luke Biewald CrowdFlower  The AI Conf, SF (2017-‐09-‐17) breakthroughs lag invention of methods;  must wait for “killer data set” to emerge,   often a decade or more 11
Design pattern: Weak supervision Creating large training data sets quickly  Alex Ratner Stanford  O’Reilly Data Show (2017-‐06-‐08) Snorkel: using weak supervision and   data programming as another instance   of human-‐in-‐the-‐loop  github.com/HazyResearch/snorkel conferences.oreilly.com/strata/strata-‐ny/public/ schedule/detail/61849 12
Design pattern: Human-‐in-‐the-‐loop Paul English on Lola's Debut for Business Travelers  Elizabeth West  Business Travel News (2017-‐10-‐04) founded 2015 by Paul English and other Kayak execs:   on-‐demand, personal travel service; uses expert travel agents for HITL initially criticized by travel industry as “competing against Siri”;   currently displacing OTAs in a reversal of “AI vs. jobs” can book on Airbnb, Southwest, etc., which aren’t available via OTA,   because of the human delegation “The first time you use Lola it’s going to be great because it’s a conversation.   We’re not making you think like a computer” “Instead of showing you 300 choices or 1,000 choices, we think we can   show you three choices, kind of good, better, best” 13
Design pattern: Human-‐in-‐the-‐loop Anand Kulkarni Crowdbotics HITL for code+test gen, trained from GitHub, StackOverflow, etc., with JIRA tickets as the granular object in the system parse specs from JIRA history, reuse what’s been done before; generate PRs for popular web stacks: React, Flask, Ruby, etc. resolve specs into the approach needed and time required,   where product managers get cost estimates, then on-‐demand expert programmers implement for you have the in-‐house engineers handle “radically novel” projects results: 1.5x software dev throughput 14
Design pattern: Human-‐in-‐the-‐loop Building a business that combines human experts and data science  Eric Colson StitchFix  O’Reilly Data Show (2016-‐01-‐28) “what machines can’t do are things around cognition,  things that have to do with ambient information, or  appreciation of aesthetics, or even the ability to  relate to another human”  15
Design pattern: Human-‐in-‐the-‐loop EY, Deloitte And PwC Embrace Artificial Intelligence For Tax And Accounting  Adelyn Zhou  Forbes (2017-‐11-‐14) compliance use cases in reviewing lease   accounting standards 3x more consistent and 2x efficient than   the previous humans-‐only teams break-‐even ROI within less than a year 16
Design pattern: Human-‐in-‐the-‐loop Unsupervised fuzzy labeling using deep learning to improve anomaly detection  Adam Gibson Skymind  Strata Data Conf, Singapore (2017-‐12-‐07) large-‐scale use case for telecom in Asia method: overfit variational autoencoders,   then send outliers to human analysts 17
Design pattern: Human-‐in-‐the-‐loop Strategies for integrating people and machine learning in online systems  Jason Laska Clara Labs  The AI Conf, NY (2017-‐06-‐29) establishing a two-‐sided marketplace where   machines and people compete on a spectrum   of relative expertise and capabilities    18
Design pattern: Human-‐in-‐the-‐loop Strategies for integrating people and machine learning in online systems Jason Laska The AI Conf establishing a two-‐sided marketplace where machines and people compete on a spectrum of relative   19 “the trick is to design systems from Day 1  which learn implicitly from the intelligence  which is already there” Michael Akilian Clara Labs
Design pattern: Human-‐in-‐the-‐loop Building human-‐assisted AI applications  Adam Marcus B12  O’Reilly Data Show (2016-‐08-‐25) “Humans where they’re best, machines for the rest.” Orchestra: a platform for building human-‐assisted   AI applications, e.g., create/update business websites  https://github.com/b12io/orchestra example: http://www.coloradopicked.com/ 20
Design pattern: Flash teams Expert Crowdsourcing with Flash Teams  Daniela Retelny, et al.   Stanford HCI  UIST (2014-‐10-‐05) computationally-‐guided teams of crowd experts supported by lightweight, reproducible, scalable team structures “elastic recruiting”: grow and shrink teams on demand, combine teams into larger organizations http://stanfordhci.github.io/flash-‐teams/ 21
Problem: disambiguating contexts
AI in Media ▪ content which can represented as   text can be parsed by NLP, then manipulated by available AI tooling ▪ labeled images get really interesting ▪ text or images within a context have   inherent structure ▪ representation of that kind of structure is rare in the Media vertical – so far 23
Disambiguating contexts Overlapping contexts pose hard problems in natural language understanding. That runs counter to the correlation emphasis of big data.  NLP libraries lack features for disambiguation.
Disambiguating contexts 25 Suppose someone publishes a book which uses the term `react`: are they talking about a JavaScript library, or about human behavior during interviews? Our customers ask for both. We handle lots of content about both. Disambiguating those contexts is important for good UX in personalized learning. In other words, how do machines help people   distinguish that content within search? Potentially a good case for deep learning,   except for the lack of labeled data at scale.
Active learning through Jupyter 26 Jupyter notebooks are used to manage ML   pipelines for disambiguation, where machines   and people collaborate: ▪ ML based on examples – most all of the feature engineering, model parameters, etc., has been automated ▪ https://github.com/ceteri/nbtransom ▪ based on use of nbformat, pandas, scikit-‐learn
Active learning through Jupyter 27 Jupyter notebooks are used to manage ML pipelines and people collaborate: ▪ ML based on examples – most all of the feature engineering, model parameters, etc., has been automated ▪ https://github.com/ceteri/nbtransom ▪ based on use of Jupyter notebook as… ▪ one part configuration file ▪ one part data sample ▪ one part structured log ▪ one part data visualization tool plus, subsequent data mining of these   notebooks helps augment our ontology
Active learning through Jupyter 28 ML#Pipelines Jupyter#kernel Browser SSH#tunnel
Active learning through Jupyter ▪ Notebooks allow the human experts to access the internals of a mostly automated ML pipeline, rapidly ▪ Stated another way, both the machines and the people become collaborators on shared documents ▪ Anticipates upcoming collaborative document features in JupyterLab
Active learning through Jupyter 1. Experts use notebooks to provide examples of book chapters, video segments, etc., for each key phrase that has overlapping contexts 2. Machines build ensemble ML models based on those examples, updating notebooks with model evaluation 3. Machines attempt to annotate labels for millions of pieces of content,   e.g., `AlphaGo`, `Golang`, versus a mundane use of the verb `go` 4. Disambiguation can run mostly automated, in parallel at scale –   through integration with Apache Spark 5. In cases where ensembles disagree, ML pipelines defer to human experts who make judgement calls, providing further examples 6. New examples go into training ML pipelines to build better models 7. Rinse, lather, repeat
Social Systems: collaboration with machines
Product management The History and Evolution of Product Management  Martin Eriksson  Mind the Product (2015-‐10-‐28) From PM’s origins as “Brand Men”,  on through the success arc of Hewlett-‐Packard,  on to Agile Manifesto, Lean Enterprise, etc. Formerly part of Engineering or Marketing,  PM now “taking a seat at the table” under CEOs 32
Conway’s Law How Do Committees Invent?  Melvin Conway  Datamation (1968-‐04) Organizations that create systems produce designs   which copy their own communication structures. For each level of delegation, someone’s scope of   inquiry narrows, design alternatives also narrow –   until a system is simple enough to be understood   in human terms. 33
Conway’s Law illustrated Organizational Charts  Manu Cornet Bonkers World Cognitive biases: ▪ anthropocentrism ▪ system justification In retrospect, Agile Manifesto   contains examples See related descriptions:  Destruction and Creation  John R. Boyd USAF  (1976-‐09-‐03) 34
First-‐order cybernetics Cybernetics: Or Control and Communication   in the Animal and the Machine  Norbert Wiener MIT  MIT Press (1948) early work had been about closed-‐loop control systems: homeostasis, habituation, adaptation, and other regulatory processes given a system which has input and output, a controller leveraging a negative feedback loop, and one or more observers outside of the system related to the early Macy Conferences 35
“the organism was no longer an input/output machine;  rather it was part of a loop from perception to action  and back again to perception” Paul Pangaro describing Jerry Lettvin @ MIT cybernetics
Second-‐order cybernetics 1. von Foerster: one can apply the understandings developed in cybernetics to the subject matter itself 2. presence of the observer is inevitable and may be desirable:   “What is said is said to an observer” 3. eigen functions: stable, dynamically self-‐perpetuating states that are self-‐referential: “We construct our realities” per constructivism 4. autopoiesis: a living entity exists as a network of components, recursively producing itself, realizing its boundaries; it grows and maintains itself by reference to itself 5. feedback loops represent conversations, from which the participants cannot be detached 6. an essentially ethical understanding 7. a productive interaction between theory and practice, in which each supports the other 37
Second-‐order cybernetics 1. von Foerster: one can apply the understandings developed in cybernetics 2. presence of the observer is inevitable and may be desirable: “What is said is said to an observer” 3. eigen functions: stable, dynamically self-‐perpetuating states that are self-‐referential: “We construct our realities” per 4. autopoiesis recursively producing itself, realizing its boundaries; it grows and maintains itself by reference to itself 5. feedback loops represent participants cannot be detached 6. an essentially ethical understanding 7. a productive interaction between theory and practice, in which each supports the other 38 second-‐order cybernetics lays a foundation for AI – it’s about the semantic relations of conversations within a system; quite apt for leveraging NLP, active learning, etc., when you have semi-‐structured dialog
Second-‐order cybernetics Autopoiesis and Cognition: The Realization of the Living  Humberto Maturana, Francisco Varela  Kluwer (1980 / original 1972) Understanding Computers and Cognition:   A New Foundation for Design  Terry Winograd, Fernando Flores  Intellect Books (1986) Conversations for Action and Collected Essays  Fernando Flores  Createspace (2013) 39
Second-‐order cybernetics ▪ biology informing computer science ▪ historical context of Project Cybersyn ▪ autopoiesis and cognition ▪ organizational closure:   “self-‐making means stability” ▪ speech acts (e.g., social analysis of open source) ▪ IMO, blueprints for AI systems Also, the focus on “information as a collection of facts”   is yet another form of cognitive bias – instilled through   30+ years of data warehouse practices, where data must   fit into dimensions, facts, schema 40
Active Learning: theory, practices, community
HITL theory: choosing what to learn Active Learning Literature Survey  Burr Settles UW Madison  (2010-‐01-‐26) Can machines learn more economically if they ask human “oracles” questions? e.g., task in-‐house experts with the edge cases? ▪ uncertainty sampling: query about instances which ML is least certain how to label -‐ least confidence / margin / entropy ▪ query-‐by-‐committee: ensemble of ML models votes; query the instance about which they disagree most ▪ expected error reduction: maximize the expected information gain of the query ▪ variance reduction: minimize future generalization error of the model (e.g., loss function) ▪ density-‐weighted methods: instances which are both uncertain and “representative” of the underlying distribution 42
HITL practices: emerging themes while ML was mostly about generalization,   now we can borrow from Frank Knight (1921):   using ML models to explore uncertainty in relationship to profit vs. risk ▪ distinguish forms of uncertainty: aleatoric (noise) vs. epistemic (incomplete model) ▪ see also: meta-‐learning [1] and [2] ▪ people who aren’t ML experts should be able to train and iterate robust models using examples ▪ emphasize use of fitness functions to make decisions, in lieu of objective functions which  tend to rely on overly simplified KPIs 43
HITL practices: model interpretation explicability of ML models becomes essential,   must be intuitive for the human experts involved:   Skater, and also Anchors, SHAP, STREAK, LIME, etc.  The Building Blocks of Interpretability  Chris Olah, et al. Google Brain  Distill (2018-‐03-‐06) Challenges for Transparency  Adrian Weller  WHI (2017-‐07-‐29) The Mythos of Model Interpretability  Zachary Lipton  WHI (2016-‐03-‐06) 44
Interpreting Machine Learning Models  Wed Mar 28 | 10-‐11 am Pacific datascience.com/resources/webinars/interpreting-‐machine-‐learning-‐models live webinar: we’ll discuss the need for methods which make the process of explaining machine learning models more intuitive, and also evaluate myths about model interpretability, from both research and business perspectives. 45 Pramit Choudhary Lead Data Scientist DataScience.com Sameer Singh CS Assistant Professor UC Irvine Paco Nathan Dir, Learning Group O'Reilly Media
HITL resources: conferences, journals, etc. HILDA 2018  Workshop on Human-‐In-‐the-‐Loop Data Analytics  Co-‐located with SIGMOD 2018  June in Houston Collective Intelligence 2018  University of Zurich, Switzerland  collocated with AAAI HCOMP 2018  July in Zurich HCOMP in Slack  https://hcomp.slack.com/ Human Computation journal  http://hcjournal.org/ojs/index.php?journal=jhc 46
HITL tooling: active learning Agnostic Active Learning Without Constraints  Alina Beygelzimer, Daniel Hsu, John Langford,   Tong Zhang  NIPS (2010-‐06-‐14) The End of the Beginning of Active Learning  Daniel Hsu, John Langford  Hunch.net (2011-‐04-‐20) https://github.com/JohnLangford/vowpal_wabbit/wiki focused on cases where labeling is expensive; uses importance weighted active learning; handles “adversarial label noise” as good or better than supervised ML, wherever supervised ML works 47
HITL tooling: machine teaching Prodigy: a new tool for radically efficient machine teaching  Matthew Honnibal, Ines Montani Explosion.ai (2017) 48
Management strategy: before In general with Big Data, we were considering: ▪ DAG workflow execution –   those are typically linear ▪ data-‐driven organizations ▪ ML based on optimizing for   objective functions ▪ general considerations about   correlation vs. causation ▪ avoid “garbage in, garbage out” 49 Jarvis workflow
Management strategy: after HITL introduces circularities: ▪ deprecate linear input/output systems   as the “conventional wisdom” ▪ analogous to an OODA loop which incorporates automation/augmentation ▪ recognize multiple feedback loops   as conversations for action ▪ recognize opportunity: loops from perception (e.g., DL) to action (e.g., HITL) and back again to perception ▪ design systems to learn implicitly   from the intelligence already there ▪ hint: recognize the “verbs” being used, rather than over-‐emphasizing “nouns” 50 Experts decide about edge cases, providing examples Experts learn through Customer interactions Customers request Sales, Marketing, Service, Training Experts gain insights via Model explanations ML Models Models focus Experts (e.g., weak supervision) Organizational Learning Human Experts Examples, Actions Customers Models act on decisions when possible Customer Use Cases Models explore uncertainty when needed
Management strategy: no-‐collar workforce No-‐collar workforce: Humans and machines in one loop  Anthony Abbatiello, Tim Boehm, Jeff Schwartz, Sharon Chand  Deloitte Insights (2017-‐12-‐05) ▪ near-‐future: human workers and machines complement the other’s efforts in a single loop of productivity ▪ 2018-‐20: expect firms to embrace a “no-‐collar workforce” trend by redesigning jobs ▪ yet only ~17% ready to manage a workforce in which people, robots, and AI work side by side – largely due to cultural, tech fluency, regulatory issues ▪ e.g., what about onboarding or retiring non-‐human workers? these are no longer theoretical questions ▪ HR orgs must develop strategies and tools for recruiting, managing, and training a hybrid workforce 51
Summary: how this matters
Conference summaries, Oct 2017 part 1  PN (2017-‐10-‐10) Themes emerging in AI conferences about the impact of ML on software process, i.e., something’s afoot: 2009–ish, data science ran headlong into prod mgmt  2012-‐ish, data sci leaders moved into prod exec roles 2018-‐ish, AI apps disrupting prod mgmt  … 53 Extrapolating trends
Flywheel Effect, circa 2018 AI drives features in products and services …   which in turn drives cloud consumption …   which in turn acquires even more data …   particularly for mobile or embedded products Incumbents now lead in AI + cloud + mobile/embed:   Google, Amazon, Microsoft, IBM, Apple, Baidu, etc.
segment assets liabilities Google, Amazon, Microsoft,  IBM,  Apple,   Baidu,   etc. ▪ AI + cloud + mobile/embed,   leveraging a flywheel effect ▪ had focused business lines well   in advance to prepare large-‐scale   labeled data sets ▪ uses AI to explore uncertainty,   focusing their core expertise ▪ high capital expenses, long-‐term R&D   as hardware evolves rapidly ▪ potential vulnerabilities by automating   too much ▪ potential vulnerabilities by mistaking   first-‐order cybernetics for second-‐order < 50% ▪ HITL provides a vector to compete   against top incumbents, with many unexplored areas of opportunity ▪ facing barriers: talent gap, competing   investment priorities, security concerns ▪ verticals eroded by horizontal business   lines from top incumbents > 50% ?? ▪ struggling to recognize business use cases ▪ buried in tech debt from digital infrastructure ▪ lacks management support Challenge: adoption by industry segment 55
What is changing and why? Second-‐order cybernetics began partly as a study of how complex systems fail, and also about what social systems   and physical systems had in common It provides foundations for AI systems of people + machines Feedback loops represent structured conversations for action, from which the participants cannot be detached The organization is no longer viewed as an input/output machine; rather it’s a pluralistic network of loops from perception to action and back again to perception –   e.g., DL augments perception and RL augments actions 56
Second-‐order cybernetics began partly as a study of how complex systems and physical systems had in common It provides foundations for Feedback loops represent structured action The organization is no longer viewed as an input/output machine; rather it’s a pluralistic network of loops from perception to action and back again to perception e.g., DL augments What is changing and why? 57 In other words, as the flywheel effect itself   is evolving, to stay ahead we must recognize the emerging “verbs”, which are entry points into the business use cases
What do organizations carry into AI? Assess the cognitive biases we bring into AI systems of people + machines: ▪ anthropocentrism and system justification, as shown by Conway’s Law ▪ DW + BI cultural lens overemphasizes “information as a collection of facts”,   missing the conversations for action ▪ digitalization sequence “Product”, “Service”, “Data”: overreacting to the nouns (facts), while ignoring the verbs (relations) ▪ delegation + committee: narrowing the scope of inquiry and design alternatives   until a system is simple enough to understand in human terms ▪ some incumbents hold tenaciously to ML apps within first-‐order cybernetics,   i.e., bias toward mostly top-‐down command and control Instead, we must design systems that learn implicitly from the intelligence already within an organization and its relationships with the customers, channels, etc. Sales, Customer Support, Professional Services, Marketing 58
What do organizations carry into AI? Assess the ▪ anthropocentrism ▪ DW + BI cultural lens overemphasizes “ missing the conversations for action ▪ digitalization sequence “Product”, “Service”, “Data”: (facts), while ▪ delegation until a system is simple enough to understand in human terms ▪ some incumbents hold tenaciously to ML apps within first-‐order cybernetics, i.e., bias toward Instead, we must design systems that learn implicitly from the intelligence already within an organization and its relationships with the customers, channels, etc. 59 Could we be encountering early stages of not-‐only-‐human cognition attempting to optimize beyond human predispositions and cognitive biases? [ed note: say at least one strange thing]
“The future belongs to those who  understand at a very deep level how  to combine their unique expertise  with what algorithms do best.” – Pedro Domingos, The Master Algorithm
The AI Conf CN Apr 10-‐13  NY, Apr 29-‐May 2  SF, Sep 4-‐7  UK, Oct 8-‐11 Strata Data UK, May 21-‐24  NY, Sep 11-‐14  SF, Mar 26-‐28 JupyterCon + events BOS, Mar 21  ATL, Mar 31  DC, May 15  NY, Aug 21-‐25 OSCON PDX, Jul 16-‐19
Get Started with NLP in Python Just Enough Math Building Data Science Teams Hylbert-‐Speys How Do You Learn? arycles, online courses, conference summaries… liber118.com/pxn/  @pacoid

Check these out next

Human in the loop: a design pattern for managing teams working with ML

Recommandé

Contenu connexe

Présentations pour vous(20)

Similaire à Human in the loop: a design pattern for managing teams working with ML(20)

Plus de Paco Nathan(20)

Dernier(20)

Human in the loop: a design pattern for managing teams working with ML