NeuroIPS () is the largest conference in the world on machine learning and artificial intelligence and the main event in the world of deep learning.
Will we, DS-engineers, also master biology, linguistics, psychology in the new decade? We'll tell you in our review.
This year the conference brought together more than 13500 people from 80 countries in Vancouver (Canada). Sberbank has been representing Russia at the conference for several years — the DS team spoke about the introduction of ML into banking processes, about the ML competition, and about the capabilities of the Sberbank DS platform. What were the main trends of 2019 in the ML community? Conference participants say: и .
Over 1400 papers were accepted at NeurIPS this year - algorithms, new models, and new applications to new data.
Contents:
- Trends
-
- Model Interpretability
- Multidisciplinarity
- reasoning
- RL
- GAN
- Basic Invited Talks
-
- “Social Intelligence”, Blaise Aguera y Arcas (Google)
- “Veridical Data Science”, Bin Yu (Berkeley)
- “Human Behavior Modeling with Machine Learning: Opportunities and Challenges”, Nuria M Oliver, Albert Ali Salah
- “From System 1 to System 2 Deep Learning”, Yoshua Bengio
Trends 2019 of the year
1. Interpretability of models and new ML methodology
The main topic of the conference is interpretation and evidence of why we get certain results. You can talk for a long time about the philosophical importance of the interpretation of the "black box", but there were more real methods and technical developments in this area.
The methodology of reproducibility of models and extraction of knowledge from them is a new toolkit of science. Models can serve as a tool for obtaining new knowledge and testing it, and every step of preprocessing, training and application of the model must be reproducible.
A significant proportion of publications is devoted not to the construction of models and tools, but to the problems of ensuring security, transparency and verifiability of results. In particular, a separate stream about attacks on the model (adversarial attacks) has appeared, and both attacks on learning and attacks on application are considered.
Article:
- — a policy article on the methodology of model verification. Includes an overview of modern tools for interpreting models, in particular, the use of attention and obtaining feature importance due to the “distillation” of a neural network with linear models.
- Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, Jonathan K. Su
- Sarah Hooker, Dumitru Erhan, Pieter-Jan Kindermans, Been Kim
- Alexander Mott, Daniel Zoran, Mike Chrzanowski, Daan Wierstra, Danilo Jimenez Rezende
- Xiao Li, Yu Wang, Sumanta Basu, Karl Kumbier, Bin Yu
- Jaemin Yoo, Minyong Cho, Taebum Kim, U Kang
- Edward Raff
ExBert.net Shows Interpretation of Models for Word Processing Tasks
2. Multidisciplinarity
To ensure reliable verification and develop mechanisms for checking and replenishing knowledge, specialists in related fields are needed who simultaneously have competencies in ML and in the subject area (medicine, linguistics, neuroscience, education, etc.). Of particular note is the more significant presence of works and presentations in neurosciences and cognitive sciences - there is a convergence of specialists and borrowing of ideas.
In addition to this rapprochement, multidisciplinarity is outlined in the joint processing of information from various sources: text and photo, text and games, graph databases + text and photos.
Article:
- Neurobiology + ML -
- VisualQA -
- RL + NLP -
Two models - strategist and performer - based on RL and NLP play online strategy
3. Reasoning
The strengthening of artificial intelligence is a movement towards self-learning systems, “conscious”, reasoning and reasoning (reasoning). In particular, causal inference and commonsense reasoning are developing. Part of the reports is devoted to meta-learning (how to learn to learn) and the combination of DL technologies with logic of the 1st and 2nd order - the term Artificial General Intelligence (AGI) is becoming a common term in the speeches of the speakers.
Article:
- Weijiang Yu, Jingwen Zhou, Weihao Yu, Xiaodan Liang, Nong Xiao
- Wang-Zhou Dai, Qiuling Xu, Yang Yu, Zhi-Hua Zhou
- Vaishak Belle, Brendan Juba
- Anton Bakhtin, Laurens van der Maaten, Justin Johnson, Laura Gustafson, Ross Girshick
- Dinesh Garg, Shajith Ikbal, Santosh K. Srivastava, Harit Vishwakarma, Hima Karanam, L Venkata Subramaniam
4.Reinforcement Learning
Most of the work continues to develop the traditional areas of RL - DOTA2, Starcraft, combining architectures with computer vision, NLP, graph databases.
A separate day of the conference was devoted to the RL workshop, which presented the architecture of the Optimistic Actor Critic Model, which surpasses all previous ones, in particular Soft Actor Critic.
Article:
- ; Kamil Ciosek, Quan Vuong, Robert Loftin, Katja Hofmann
- ; Yasuhiro Fujita (Preferred Networks, Inc.)*; Toshiki Kataoka (Preferred Networks, Inc.); Prabhat Nagarajan (Preferred Networks); Takahiro Ishikawa (The University of Tokyo) [external pdf link].
- ; Danijar Hafner (Google)*; Timothy Lillicrap (DeepMind); Jimmy Ba (University of Toronto); Mohammad Norouzi (Google Brain)
StarCraft players fighting with the Alphastar model (DeepMind)
5.GAN
Generative networks are still in the spotlight: a lot of work uses vanilla GANs for mathematical proofs, as well as applying them in new, unusual ways (graph generative models, working with series, applying to causal relationships in data, etc.).
Article:
- Sangwoo Mo, Chiheon Kim, Sungwoong Kim, Minsu Cho, Jinwoo Shin
- Dan Zhang, Anna Khoreva
- Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, Kalyan Veeramachaneni
Since the work was accepted more Below we will highlight the most important performances.
Guest Talks
“Social Intelligence”, Blaise Aguera y Arcas (Google)
The report is devoted to the general methodology of machine learning and the prospects that are changing the industry right now - what crossroads are we facing? How does the brain and evolution work, and why do we make so little use of what we already know well about the development of natural systems?
The industrial development of ML largely coincides with the milestones in the development of Google, which publishes its research on NeurIPS year after year:
- 1997 - launch of search facilities, first servers, low computing power
- 2010 - Jeff Dean launches the Google Brain project, the neural network boom at the very beginning
- 2015 - industrial implementation of neural networks, fast face recognition directly on the local device, low-level processors, sharpened for tensor computing - TPU. Google launches Coral ai - an analogue of raspberry pi, a mini-computer for introducing neural networks into experimental setups
- 2017 - Google starts developing decentralized learning and combining neural network training results from different devices into one model - on android
Today, an entire industry is dedicated to data security, aggregation and replication of learning outcomes on local devices.
- the direction of ML, in which individual models learn independently from each other, and then are combined into a single model (without centralization of the source data), adjusted for rare events, anomalies, personalization, etc. All Android devices are essentially a single computing supercomputer for Google.
Generative models based on federated learning is a promising future direction according to Google, which is “in the early stages of exponential growth.” GANs, according to the lecturer, are able to learn how to reproduce the mass behavior of populations of living organisms, thinking algorithms.
Using the example of two simple GAN architectures, it is shown that in them the search for an optimization path wanders in a circle, which means that optimization does not occur as such. At the same time, these models very successfully simulate the experiments that biologists put on populations of bacteria, forcing them to learn new strategies for behavior in search of food. It can be concluded that life works differently than the optimization function.
Wandering GAN optimization
Everything that we do in the framework of machine learning now is narrow and highly formalized tasks, while these formalisms are poorly generalized and do not correspond to our subject knowledge in such areas as neurophysiology and biology.
What is really worth borrowing from the field of neurophysiology in the near future is new neuron architectures and a little revision of the backpropagation mechanisms.
The human brain itself learns not like a neural network:
- He does not have random primary inputs, including those laid down through the senses and in childhood.
- He has laid down directions of instinctive development (the desire to learn a language from an infant, upright posture)
Training an individual brain is a low-level task, perhaps we should consider "colonies" of rapidly changing individuals passing knowledge to each other in order to reproduce the mechanisms of group evolution.
What can we adopt in ML algorithms now:
- Apply cell lineage models that provide learning to the population, but short life of the individual (“individual brain”)
- Few-shot learning on a small number of examples
- More complex neuron structures, slightly different activation functions
- Transferring the “Genome” to the Next Generations – Backpropagation Algorithm
- Once we combine neurophysiology and neural networks, we will learn how to build a multifunctional brain from many components.
From this point of view, the practice of SOTA solutions is detrimental and should be reviewed in favor of the development of common tasks (benchmarks).
“Veridical Data Science”, Bin Yu (Berkeley)
The report is devoted to the problem of interpretation of machine learning models and the methodology of their direct testing and verification. Any trained ML model can be perceived as a source of knowledge that needs to be extracted from it.
In many areas, especially in medicine, the application of the model is impossible without extracting this hidden knowledge and interpreting the results of the model - otherwise we will not be sure that the results will be stable, non-random, reliable, and will not kill the patient. A whole area of work methodology is developing within the deep learning paradigm and goes beyond it - veridical data science. What it is?
We want to achieve such a quality of scientific publications and reproducibility of models that they are:
- predictable
- computable
- stable
These three principles form the basis of the new methodology. How can ML models be checked against these criteria? The easiest way is to build immediately interpretable models (regressions, decision trees). However, we also want to get the immediate benefits of deep learning.
Several existing ways to work with the problem:
- interpret the model;
- use methods based on attention;
- use ensembles of algorithms when learning, and ensure that linear interpretable models learn to predict the same answers as a neural network, interpreting features from a linear model;
- change and augment training data. This includes the addition of noise, interference, and data augmentation;
- any methods that make sure that the results of the model are not random and do not depend on small unwanted interference (adversarial attacks);
- interpret the model after the fact, after training;
- study feature weights in various ways;
- study the probabilities of all hypotheses, the distribution of classes.
Adversarial attack
Modeling errors are costly to everyone: a vivid example is the work of Reinhart and Rogov ""influenced the economic policy of many European countries and forced them to pursue a policy of savings, but a careful recheck of the data and their processing years later showed the opposite result!
Any ML technology has its own life cycle from implementation to implementation. The goal of the new methodology is to test on three main principles at each stage of the model's life.
Results:
- Several projects are being developed that will help the ML model to be more reliable. This is, for example, deeptune (link to: );
- For further development of the methodology, it is necessary to significantly improve the quality of publications in the field of ML;
- Machine learning needs leaders with multidisciplinary training and expertise in both technical and human fields.
“Human Behavior Modeling with Machine Learning: Opportunities and Challenges” Nuria M Oliver, Albert Ali Salah
Lecture dedicated to the modeling of human behavior, its technological foundations and prospects for applications.
Human behavior modeling can be divided into:
- individual behavior
- small group behavior
- mass behavior
Each of these types is amenable to modeling using ML, but with completely different input information and features. Each type also has its own ethical issues that each project goes through:
- individual behavior - identity theft, deepfake;
- the behavior of groups of people - deanonymization, obtaining information about movements, phone calls, etc.;
individual behavior
To a greater extent, it concerns the topic of Computer Vision - recognition of human emotions, his reactions. It is possible only in context, in time, or with the relative scale of his own variability of emotions. On the slide - recognizing Mona Lisa's emotions with the help of context from the emotional spectrum of Mediterranean women. The result: a smile of joy, but with contempt and disgust. The reason is most likely in the technical way of defining a “neutral” emotion.
Behavior of a small group of people
So far, it is the worst modeled due to insufficient information. As an example, the works of 2018-2019 were shown. on dozens of people X dozens of videos (cf. 100k++ image datasets). For the best modeling within the framework of this task, multimodal information is required, preferably from sensors on a body-altimeter, thermometer, recording from a microphone, etc.
Bulk Behavior
The most developed direction, as the UN and many states act as the customer. Surveillance cameras, telephone tower data - billing, SMS, calls, data on movement between state borders - all this gives a very reliable picture of the movement of people flows, of social instabilities. Potential applications of the technology: optimization of rescue operations, provision of assistance and timely evacuation of the population in case of emergency. The models used are mostly poorly interpreted so far - these are various LSTMs and convolutional networks. There was a brief remark that the UN was lobbying for a new law that would oblige European businesses to share anonymized data necessary for any research.
“From System 1 to System 2 Deep Learning”, Yoshua Bengio
In a lecture by Joshua Bengio, deep learning meets neuroscience at the level of goal setting.
Bengio distinguishes two main types of problems according to the methodology of Nobel laureate Daniel Kahneman (book "»)
type 1 - System 1, unconscious actions that we do "on the machine" (ancient brain): driving a car in familiar places, walking, recognizing faces.
type 2 - System 2, conscious actions (cerebral cortex), goal setting, analysis, thinking, compound tasks.
So far, AI has reached sufficient heights only in tasks of the first type, while our task is to bring it to the second, by teaching it to perform multidisciplinary operations and operate with logic, high-level cognitive skills.
To achieve this goal, it is proposed:
- use attention as a key mechanism for modeling thinking in NLP tasks
- use meta-learning and representation learning to better model features that affect consciousness and their localization - and on their basis move on to operating with higher-level concepts.
Instead of a conclusion, we leave an invited talk entry: Bengio is one of many scientists who are trying to expand the field of ML beyond optimization problems, SOTA and new architectures.
The question remains open to what extent the combination of the problems of consciousness, the influence of language on thinking, neuroscience and algorithms is what awaits us in the future and will allow us to move on to machines that “think” like people.
Thank you!
Source: habr.com