AAAI 2019 Notes.pdfAAAI 2019 Notes， AAAI大会是国际人工智能领

文件名称: AAAI 2019 Notes.pdf

所属分类: 机器学习

开发工具:

文件大小: 632kb

下载次数: 0

上传时间: 2019-09-01

提供者: ccy***

下载 (632kb)

不能下载？报告错误

详细说明：AAAI 2019 Notes， AAAI大会是国际人工智能领域的顶级国际会议。Friday February 1st 67 7. 1 Rcinforccmcnt Learning 67 7.1.1 Diversity-Driven Hierarchical RL 36 67 7.1.2 Towards Better Interpretability in DQN 7.1.3 On RL for Full-Length Game of Starcraft 28 69 7.2 ReasOning under Uncertainly 70 2.1 Collecting Onlinc Lcarning of CPs in Multi-Agcnt Systcms 70 7.2.2 Weighted Model Ingeration using Know ledge Compilation Off-Policy Deep RL by Bootstrapping the Covariate Shift .72 7.2.4 Compiling Bayes Net Classifiers into Decision Graphs 35 7 This document contains noles I took during the events I Managed to nake it to aL AAAI in Honolulu, Hawaii, USA, including sessions of thc Doctoral Consortium. Plcasc fccl frcc to distributo it and shoot me an email at david-abelobrown edu if you find any typos or other items that need correcting. 1 Conference Highlights AAAi was fantastic- the invited talks offered impressive videos, inspiring visions of the future and excellent coverage of many areas, spanning game playing, learning, human-robot interaction data management, and exciting applications. I also enjoyed the two evening events: 1)the 20 year roadmap for AI research in the US, and 2)the debate on the future of Al. Both events raised compelling questions for researchers and practitioners of Al alike I also want to highlight the doctoral consortium(DC). This was the first DC I have participated in; in short, I strongly encourage grad students to do at least one dc during their program. You will get exposure to some fantastic work being done by peers all around the world and receive tailored mentorship on your presentation skills, how you write, and your research objectives and methods more generally AAAI really struck me as doing a great job of mixing together many subfields that don't often spend as much time talking to one another-I met plenty of folks working in planning, constraint satisfication, automated theorem proving. Al and society, and lots of ML/RL researchers A final point that was raised at the roadmap- naturally, a huge fraction of research /industry is concentrated on ML at the moment. But, it's important that we continue to push the frontiers of knowledge forward across many different areas. So, if you re considering going into grad school soon, do consider pursuing other topics /areas that offer fundamental and important questions which there are many! beyond ML And that’ s that!Let’ s dive in. 2 Sunday january 27th: Doctoral consortium It begins! Today I'll be at the Doctoral Consortium(DC)my goal with the notes is both to give folks a sense of what a DC entails, and to share the exciting research of some great grad students 2.1 Overview of the dc I highly recommend doing a doctoral consortium at some point during grad school. I learned a huge amount from the experience For those that dont know, a dC involves preparing a short abstract summarizing your work, and iving a 10-20 minute presentation to your peers and their montors. Each student participating is assigned a. mentor(from their area. )that helps with preparing your presentation and gives you more general advice on your research It was a great experience! I had the pleasure of meeting many wonderful grad students and hearing about their work 2.2 Neeti Pokhriyal: Multi-View Learning From Disparate Sources for Poverty Mapping Focus: Learning from multiple disparate data sources, applied to sustainability and biometrics Specific Application: Povert mapping. Spatial representation of economic deprivations for a coun try. A major tool for policy planners Current method is a household survey, which is 1) costly, 2)time consuming, 3 only available for small samples Research Goal: Get accurate, spatially detailed and diagnostic poverty maps for a country Lots of data availablc via weather, strcct maps, cconomic data, mobilc phones, satellite imagery But! each of these data sources are structured very different Definition 1(Multi-View Learning): A stgle of learning takes as input separate, semantically distinct kinds of data, and brings them together into a factorized representation for use in predictive models Method: learn a Gaussian Process(GP)Regression model combined with elastic net regularia- tion图S. Using this model yields the map pictured in Figure I Then perform quantitive analysis and vali dates that their model is making high qua ity predictions by comparing Objective 2: learn a factorized representation from multiple data sources. The hope is that we can disentangle explanatory factors that are unique to each data source Poverty Map of Senegal africa r bern p the moolai abrsm hl ea ni wrol commune Figure 1: Higher fidelity poverty prediction Sort of an EM like approach: 1. Learning Step MAp views y and z to shared subspaces xi Inference Step: Pcrform infcrcncc on thcsc subspaccs Q: Main question, then: how do we learn the shared subspace? A: Separate data belonging to different class across different views is maximized, while ensuring alignment of projects from each view to the shared space. Can be solved using a generalized Eigenvalue problem, or using the kernel trick 2.3 Negar Hassanpour: Counterfactual Reasoning for Causal Effect Estimation Problem: Consider Mr. Smith, who has a disease and some known properties(age, bMi, etc.) Doctor provides treatment X and observes the effect of treatment X (but does not get data about the counterfactu: what would have happened if doc had applied treatment Y? Goal: Estimate the "Individual Treatment Effects"(ITE)-how does treatment X compare to Y? Datasets Randomized Controlled Trial(RCT): See lots of both X and Y. But, it's expensive (lots of trials) and unethical(giving placebos when you know the right treatment) Observational Study: provide the preferred treatment. But, sample selection bias F ample: Trea. ting heart, disease, a doc prescribes surgery to younger patients and medication to older patients. Compare survival time- but, clear bias in who gets what treatment This is a really fundamental problein called"sample selection bias"- rich palients receiving pensive trcatmcnt vs. poor paticnts rccciving cheap trcatment and so on Overview of this work Generate realistic synthetic datasets for evaluating these methods(since good data is hard to come by ->Take an rct and augment it with synthetic data e Use representation learning to reduce sample selection bias Want Pr((a)l=0)N Pr(o() (=1) lo be sinilar, with c the learned representation and t the treatment o Learn underlying causal mechanism with generative models c Learn causal relationships between treatments and outcomes by using generative models an we identify the latent sources of outcome froin observational dataset? e Perform survival predictions Can we predict outcomes that are censored or take place after studies end? e Going beyond binary treatments > Many, but not all, treatments are binary. Can we go beyond this to categorical or real valued treatments? Providing a course of treatment > Call on reinforcement learning 2.4 Khimya Khetarpal: Learning Temporal abstraction Across Action per- ception Q How should an Ai agent efficiently represent, learn, and use knowledge of the world? A: Lets use temporal abstractions xample: preparing breakfast. Lots of subtasks/activities involved like(high level): choose eggs lype of toast(Imid level)chop vegetables, get butler, and (low level)wrist and arIn Movements Definition 2(Options 37 ) An option formalizes a skill/temporally c tended action as a triple: (, 1, B, r, wwhere Ic S is a initiation set, B: S- Pr(S)is a term..ation probability and:S→> A is a policy. Example: A robol navigates Through a house between two rools. To do so. it has to open a door We lct I denotc the states whcrc the door is closcd, 6 is 1 when the door is opcn and 0 otherwise and T opens the door. Then, this option defines the "open the door"skill Main Question: Can we learn useful temporal abstractions? Hypothesis: Lcarning options which arc specialized in situations of spccific intcrcst can bc uscd to get the right temporal abstractions Motivation: Al agents should be able to learn and develop skills continually, hierarchically, and incrementally over line So, imagine we had a house decomposed into different rooms. Then we would like to learn skills that take the agent between each room. Further, the agent should be able to transfer for one agent to another Objective 1: Lcarn options and intcrcst functions simultancously New idea: break the option-critic assumption 2 that I=S. Instead, consider an interest func- tion: Definition 3(Interest Function An interest function is an indication of the ectent to which an option is interested in state s Now learn a policy over options and an interest function- we can jointly optimized over both things Derive the policy gradient, theorem for interest functions, intra-option policy, and the termination function Theme I: Learning options with interest functions 日回图團围围 Figure 2: Learned interest functions Also explore learning interest functions in continuous control tasks, showing nice separation be tween the learn options Objective 2: Considcr a ncvcr-cnding strcam of pcrccptual data. Wc'd like to learn a strcam of percepts and behavior over time Challenges How can we the agent automatically learn features which are meaningful pseudo rewards? Where to task descriptions come from? How can we achieve the most general options without hand designing tasks/rewards Evaluation in a lifelong learning task? Benchmarks? 2.5 Ana Valeria Gonzalez-Garduo: RL for Low Resource Dialogue Systems Goal 1: Create more informed approaches to dialogue generation Goal 2: Use Rl for domain adaption in goal oriented dialogue And: can we do this in a language agnostic way? So, introduce models that can work with any/many languages) Dialogue svstems are divided into two subfields 1. Open ended dialogue generation: typically use encoder-decoder architectures 2. Goal or'ienled dialogue: predominantly tackle using "pipeline"nethods. So, automatic speech rocognition unit, then an undcrstanding unit, and so on Current Focus: state tracking". That is, state tracking deals with inferring the user intent or belief state during the conversation But, limitation: intents usua lly rely on a particular ontology that defines which intents are valid Current Status of the Project: Bridge the gap inb goal oriented dialogue. Main goal: can we get rid of the need for annotations? General idea given a. bot s utterance("how can I help? "), and a user response("I want to change payment methods), we want to find a relevant query from prior conversations to identify what the user said. Or really, use it to condition the decoder Result: this model works very well! On bleu their model performs favorably, but more impor- tantly, on a human evaluation, their responses were consistently chosen over the baseline Q: But, what if our domain is not in the pool of relevant conversations A: Work in progress! Idea - Use RL 1. Phase 1: Use existing dala for slate tracking, pretrain Inodels in a supervised manner Turn level supervision, slots and values represented using word embeddings 2. Phase 2: Use RL to finetune pretrained model Rely on dialogue level supervision (joint goal accuracy) as reward. So, how many slot values("Food-Mexican, Price-Cheap), to determine the reward Challenges in using RL for state tracking: dialogue is long(credit assignment is hard! ) sample efficiency, might be able to leverage curriculum learning Main Future Direction: Enable dialogue stale transition Inodel lo generale new unseen slots 6 AAAI Tutorial: Eugene Freuder on How to Give a talk Start with an example! Or a counter example These arc just his conclusions! So decide for yourself, of coursc This talk is not intended to be mean spirited -he'll be talking about mistakes people make Meta-message: presenting a talk is a skill that can be studied and practiced! And it's worth doing spend years researching and 10 minutes presenting. The 10 minutes should be polished Six points Convey enthusiasm 2. Make it Easy to follow 3. Employ examples 4. Expressive 5. Enhance your presentation with visuals/dynamic material 6. Engage the audience 2.6.1 Enthusiasm The secret of a good talk: Enthusiasm! If you're not enthusiastic about your work, how do you expect anyone else to be? Fcar of public speaking: glausophobia-ranked as thc most common fcar in the USa(morc so than spiders / death)

(系统自动生成,下载前可以参看下载内容)