The International Conference on Machine Learning (ICML) is one of the leading international academic conference in machine learning. The 35th event took place in Stockholm, Sweden, between July 10th and July 15th 2018. Around 600 research papers were presented to more than 5,000 attendees. Speakers belonged to cutting-edge R&D teams from prestigious universities (Berkeley, Stanford, EPFL, Tsinghua University…) and business corporations (Google, Amazon, Netflix…).
ICML unveils the latest breakthroughs and provides insights on future research in Machine Learning. It also makes networking and collaboration easier as a great part of the research community is gathered at the same place. Here are our main take-aways !
It is now quite clear that Deep Learning has reached some kind of maturity. The AI community no longer discusses its scope of opportunities, but focuses on its social consequences: how to make it greener, more ethical and more secure.
On the ecological side, Max Welling, research chair in Machine Learning at the University of Amsterdam and VP Technologies at Qualcomm, advocates that models should be evaluated in terms of “intelligence per kilowatt-hour”. As the complexity of Deep Learning algorithms is constantly growing, it becomes standard to train huge algorithms for each task, even if the use case does not need this complexity. Consequently, the cost of storing such models, in terms of electricity and hardware consumption is vainly increasing. To help solving this issue, Max Welling promotes a selective sorting of parameters, as we can do in our every-day life with our wastes in order to reduce our ecological impact.
As ML is more and more affecting our everyday lives, algorithm fairness is also becoming a crucial research field. Indeed, unintended discrimination on race, religion or gender can frequently arise because of biases that shaped training datasets. “Delayed Impact of Fair Machine Learning” addresses this issue and was given a best paper award. This work proves that classical fairness criteria, applied on snapshot of training populations, do not ensure fairness over time. It also describes a simple retroactive fairness criterion, that can sustainably affect a population towards legally and morally fairer decisions.
Deep learning models also need to be more secure, especially from adversarial examples. An adversarial example is an algorithm input intentionally designed to cause the model to make a mistake. Let us take an example: an autonomous car needs to recognize road signs. On the image below, the stop sign has been slightly modified from its original consistency. Any human could easily determine that this road sign is a “STOP”, however if the algorithm is not robust enough to adversarial examples, this stop sign could be misclassified. This example is quite striking as a misclassification won’t make the autonomous car stop, which would endanger human lives!
Reinforcement Learning (RL) has become very popular since Google DeepMind algorithm AlphaGo defeated the world champion Go player Lee Sedol in 2017. RL aims at training a software agent (AlphaGo) that makes some actions (putting stones) in an environment (a go board) in order to maximize some notion of cumulative reward (winning the game). RL proved to perform a lot better in very complex situations, when computing an exact solution with classical methods such as Markovian Decision Processes would be too computationally expensive.
In RL, the main dilemma is to find the right trade-off between exploitation and exploration: exploitation means choosing a known action that brings a guaranteed reward, while exploration means trying unknown actions hoping to find a more rewarding policy. Some algorithms already prove good performances in finding optimal policies, under some predefined circumstances, as the Q-Learning algorithm.
A lot of papers at ICML dealt with generalization to several software agents, several tasks, or finding an optimal policy in complex environments where very few actions are rewarded. Another area for research deals with RL interpretability: the paper “Visualizing and Understanding Atari Agents” aims at producing easily understandable policies, unlike deep RL where policies are more like “black boxes”.
There are plenty of research themes and applications in this field! The new astonishing challenge for Google Deep Mind is to train a software agent able to play Starcraft 2. This game has been chosen for its huge amount of different possible configurations, that forces the agent to be very reactive.
Other trendy algorithms are the Generative Adversarial Networks (GAN), which are mainly used to generate realistic artificial images. Many applications exist for GANs such as adversarial example generation (the STOP sign we mentioned earlier), face aging, anime character creation or even text to image conversion. One striking use case is domain transfers: it consists in mapping an input image to a new image with different aspect or texture learned from the training dataset. On the image below, the input sample is only the edges of a bag. As we can see on the right picture, the GAN is able to output a realistic image of the bag by enriching the initial sketch.
These algorithms involve two neural networks competing against each other to train a performant model. The first neural network is the generative one, it produces new artificial images, inspired from the original dataset. The second network is the discriminative one, it tries to classify the images as real or artificial. The aim of the generative network is to produce images that could not be detected as artificial pictures by the discriminative network. These antagonist networks constantly challenge themselves to produce impressive results.
One of the biggest issues for research with GANs is quality measures, that is to say, how GANs performances can be compared one to the other. We know that these models perform well because we can observe it, but how can we formally define what a realistic image is? Yet, researchers are currently working on evaluation methods based on geometrical properties of the datasets, as in the paper “Geometry Score: A Method for Comparing Generative Adversarial Networks”. The authors suggest to compare the generated dataset of artificial images to the training dataset through some geometrical considerations, to assess their similarities.
Besides, the lack of universal evaluation method illustrates that GANs are still an early stage concept: this means they will keep providing areas for research and impressive applications for some time. As an example, the paper “eCommerceGAN : A Generative Adversarial Network for E-commerce” suggests an appealing application of GANs for the Retail Industry. It uses GANs to generate plausible customer transactions. Such tool could help understanding customer behavior, or anticipating customer reactions to the introduction of new products.
He graduated from CentraleSupelec and got a Master Degree in Machine Learning
He previously worked on R&D projects at Groupe Salins and Université Pierre-et-Marie-Curie.
He contributes to the conception and implementation of Machine Learning algorithms at UntieNots since April 2017.