Upcoming seminars and events
Our reading group meets quaterly and provides foundations and up-to-date information on topics in power-efficient deep learning, mathematical statistics and optimization. We give participants valuable experience in leading group discussions and share state-of-the-art machine learning. For each session, papers are assigned in advance, and one to two participant guide the discussion.
Click here to the previous reading group !
Past seminars and events
• July, 3rd 2023 - Florine Lefer, Team member.
Title: From soil to plate : A Comparative Study of Vegetable Quality
This presentation introduces the project "From Soil to Plate." The project, funded by EMAC (64), aims to compare the sensory quality of vegetables derived from various agricultural practices and market types, such as conventional, organic, no-till, and direct selling. Agricultural practices pose significant concerns for humanity, including pollution, soil erosion and degradation, increasing food demand, and a decline in nutritional values of products. It is for these reasons that our team has taken on this subject, conducting nutritional and sensory analyses of vegetables from diverse agricultural practices to address the following questions: Are there any differences? Do people have preferences? Do virtuous practices result in better vegetable quality? The presentation provides an overview of the current state of the art in sensory and nutritional analysis of vegetables from different agricultural practices. It also delves into the project itself, outlining its methodologies. Lastly, we will present our initial findings.Slides here
• June, 12th 2023 - Paul Gay, Team member.
Title: Integrating environmental impact of AI in a data center
The growing environmental impact of our digital practices is partly due to the development of technologies such as artificial intelligence. Thus, it seems relevant to measure the footprint of these algorithms, at the source, in the data centers in order to benefit from a controlled environment allowing accurate measurements, and to raise the awareness of data scientists, who have an important role in this development. We propose to estimate the carbon footprint and assess the "environmental impact"/"algorithm accuracy" trade-off for deep learning algorithms at data center scale: algorithm iteration, user , node, entire infrastructure. The application context will be centered on the LabIA, a cluster of 12 stations dedicated to AI and used by 5 laboratories. The feedback provided to users will have two objectives: (i) detect bad practices and improve energy efficiency and user experience (ii) Increase awareness regarding the impact of their practices, both arising from the energy that their algorithms use on the cluster, and the consequences and rebound effects of the applications they develop. The data collected will make it possible to identify new opportunities applicable to other large data centers such as the Jean-Zay data center.Slides here
• May, 22th 2023 - Nicolas Tirel, Team member.
Title: Automatic Speech Recognition for children voices in MathIA solution
With the objective to summarize the work done since the beginning of the collaboration with Prof en Poche, we will go through the whole system of automatic speech recognition and we will see how we deal with children voices without any big dataset available publicly.Slides here
• April, 3th 2023 - Sébastien Loustau, Team member.
Title: Bayesian Deep Active Learning (Bayesian Active Learning)
In this talk, I will present the contribution of a new Bayesian optimizer to the problem of selecting new data for text and image classification. After an introduction of the Bayesian approach in Deep Learning which will present an optimizer based on a VI (Variationel Inference) approach, and a presentation of the datasets and the learning task (multi-label classification), we will detail the methodology of the computation of the uncertainty in the output of neural networks, allowing to obtain new entropy measures based on the randomness of the Bayesian network. Then we will compare the results of active learning where the selection of data is derived from classical approaches based on the entropy of the output of an Adam-type optimizer, or a Bayesian variational inference optimizer.Slides here
• March, 20th 2023 - Matthieu François, Team member.
Title: Artificial intelligence integration to crisis management. Part IV. Active Learning and Social Computing
In this presentation, we will be discussing our latest developments in Active Learning and uncertainty metrics applied to crisisMMD data. After showcasing our first set of results with the last seminar, we have now gone further by introducing new diversity metrics and presenting more robust results. Additionally, we will delve into the exciting field of social computing and opinion mining.Slides here
• March, 6th 2023 - Simon Lebeaud, Team member.
• February, 13th 2023 - Nicolas Tirel, Team member.
Title: Application of Data Adaptation and Domain Adaptation for Automatic Speech Recognition purposes
Explanation on the construction and the use of Data and Domain Adaptation in the case of Automatic Speech Recognition.Slides here
• January, 30th 2023 - Paul Gay, Team member.
Title: How to build French NLP models to study environmental transitions.
How can social computing help to understand social behabvior in environment transition ? An important issue for social science researchers is to identify the weight of the different viewpoints and to obtain their evolution over time. By analyzing broadcastnews and social networks, TAL technologies has the potential to provide additional information. In these condition which language model is the best suited for the task at end, and how to favor the design of new language models for social computing. This seminar will present our efforts to build a set of tasks which requires little supervision and enable to compare different models or monitor the performance of a transformer model during its training. These tasks include few shot or zero shot classification and retrieval. Results with standard NLP models (Fasttext, Flaubert, Camembert) will be presented.Slides here
• January, 9th 2023 - Matthieu Mastio
Title: Emission-reducing deployment of shared office networks
We propose a step by step method based on publicly available data to design a master plan optimizing the number of coworking spaces to deploy in a given area, as well as their placements across the territory. We aim to develop a systemic approach with the objective of reducing GHG emissions from commuting. We design and implement a decision model that describes when a person would choose to work in a shared office and which transport mode they would use. Then, we propose two methods that work on different scales and complement each other. The first one is on a macroscopic level, where we collect and preprocess population data to convert them into an exploitable form and perform an optimization of the co-working spaces placement to maximize the gains in terms of traveled distance, using a linear solver and heuristic algorithms. In the second method, we create a synthetic population and perform a mesoscopic traffic simulation of the observed territory with MATSim, before and after the addition of the co-working places. This allows us to observe how the users would adapt their activity chains and to analyze the impact on the traffic and the usage of different facilities.
• January, 5th 2023 - Mathieu Brugidou
Title: Que peuvent les plongements lexicaux pour l’analyse sociologique des textes ? Analyser les discours et caractériser les locuteurs des plateformes « Grand Débat National » et « Vrai Débat »
This seminar will present a contribution to the evaluation of the contribution of so-called "word-folding" algorithms to the sociological analysis of texts: on the one hand, by confronting the results of the semantic analyses of these algorithms to the now well-known approaches of textual data analysis or textometry; on the other hand, by focusing on what constitutes one of the main obstacles to the sociological analysis of the web: the difficulty to sociologically characterize the authors of statements coming from the web. To do this, we analyze statements from "civic tech" platforms - a governmental platform, the "Grand Débat National", and its political and algorithmic riposte proposed by a collective of Gilets jaunes, the "Vrai Débat". A third corpus from the "Entendre la France" platform, with the same design as the "Grand Débat National" and otherwise documented in terms of socio-political properties, will allow us to characterize the speakers according to their speeches and to try to predict by machine learning approaches "pseudo-properties" assigned to the speakers of the "Grand Débat National".
Philippe Suignard is an expert researcher at EDF R&D. He is interested in new technologies for customer relations (customer and advisor side), including voice processing, text mining, chatbots and social networks. His work focuses on classification (supervised or unsupervised) of textual data, lexical embeddings and their applications in business, such as email processing, complaints, oral conversations, technical reports, opinion analysis or argument detection.Google Scholar, HAL
• December, 12th 2022 - Sébastien Loustau, Team member.
Title: How to build a viable business in a context of degrowth?
In this seminar, we start with an introduction to the context of degrowth. From the macro-economical point of view, the degrowth theory proposes to build a new economical paradigm where the growth of the economical value at the macro-economical scale is not the main focus. It is based on a reject of the standard business-as-usual model, and the introduction of new - non measurable - dimension into the process: well-being, biodiversity, personal production, and social links. After this introduction, we study the gap between this global point of view and recommandations and the micro-economical problem of a small firm business model: how to build a viable firm in this new economy? For that purpose, we present a phd where a theoretical framework for business in a degrowth context is proposed. This framework is based on several markers like energy throughput, internal business operation, governance, ownership and also barriers. This framework is also tested over a panel of small firms, and apply to design the next level of our GreenAI UPPA Team in 2023!Slides here
• December, 06st 2022 - Hugo Tessier, IMT Atlantique.
Title: How to reduce the hardware cost of a neural network through pruning
Since 2015, facing the increasing cost of deep neural networks, pruning has become a very active research area, within all existing compression methods. However, the difficulty of materially exploiting the multiplication of hollow matrices has led to the emergence of different methods, aiming at achieving a more easily exploitable form of sparsity: this is called structured pruning. In this talk, we will see what are the practical and theoretical issues of structured pruning, the different methods that can be found in the literature, and the main difficulties encountered in the field, to finally reach more general considerations about pruning as a search for architectures.Slides here
• November, 21st 2022 - Matthieu François, team member.
Title: Event Detection and Uncertainty, a crisis management context
The confidence of a model in its predictions is essential for it to be robust and for its interpretation. After a brief retrospective on the integration of our MedIA tool into Cwall, a state of the art on Active Learning methods and Uncertainty Measures will be presented in a context of event detection. Motivated by a fine annotation of our data, we will then discuss unsupervised methods. Finally, we will discuss our first measurements on crisis datasets using the fine-tuning of the BLOOM model [BigScience].Slides here
• November, 14th 2022 - Fatou Kiné Sow, team member.
Title: Advances in trash detection project: presentation of YOLOv7 and methods for addressing class imbalance
In order to propose an immersive awareness solution by putting digital, AI and user experience at the service of the environmental cause, we are implementing a real-time trash detection system to raise awareness about recycling in schools. During the previous seminar on this topic, we showed the first results obtained with the object detection model YOLOv5 on our training dataset made of images of trash taken in nature. To improve these results and to add in our system the possibility to have the mask of the found object, we experimented the new version of YOLO (YOLOV7), the fastest and most accurate real time object detection model at the moment. The following seminar presents the different optimizations brought to this new version of YOLO and some methods used to address the problem of class imbalance in a dataset.Slides here
• November, 7th 2022 - Pierre Cilluffo Grimaldi, Phd at CELSA.
Title: Social construction and globalization of wilderness : international public opinion vs people's doubts in the case of Amazon rainforest.
We develop the role of environmental imaginaries (wilderness, "Eldorado", etc) for biodiversity conservation in Amazon rainforest, its interrelation with international public opinion and the resulting conflicts of representation between diverse populations ("determined territory" vs "territory prescribed"). Ultimately, it will be question of technology and biodiversity conservation. We think about the possibilities of empowerment that technology offers to reduce this dichotomy and their contributions of co-constructed and objectifying informationSlides here
• October, 17th 2022 - Simon Lebeaud, team member.
Title: A review of state of the art tracking and object detection
From calculating euclidean distance to deep appearance metrics association, state of the art tracking techniques changed quite a lot in the past decade. Evolving in parallel to object detection and relying a lot on it, tracking techniques are trying to solve the association between object states while coping with the sometimes unreliable detections.Slides here
• October, 10th 2022 - Nicolas Tirel, team member.
Title: Advances in ASR for schoolchildren
During the previous seminar, we presented an ASR for children in a classroom context to learn mathematics using their voices. We used at this time DeepSpeech, an open-source implementation developed by Mozilla and inspired by the Silicon Valley AI Lab, with a result of 18% in WER. Since that, we have changed DeepSpeech to Coqui STT, adapt the corpus with a more realistic one and designed a specific language model to get even better results. The following seminar will go through our journey, with a more spontaneous dataset and a lighter model.Slides here
• October, 4th 2022 - Samuel Rincé.
Calculating the multi-criteria impacts of the manufacture and use of digital equipment, servers, and the cloud.
Boavizta is an association dedicated to the development of methodologies, tools, and datasets that facilitate the measurement of the multi-criteria environmental impact of digital equipment. Boavizta's work is open source and state-of-the-art in terms of life cycle assessment. The group has about 100 members involved in about 15 projects. The presentation will start with the mission of Boavizta and the importance of the different works carried out within the association. We will talk in particular about two tools developed internally and open source. Datavizta is a website for visualizing carbon impact data provided by manufacturers. Boaviztapi is a multi-criteria impact calculation API for servers, cloud instances, or electronic components (CPU, RAM, etc.). We will also present the bottom-up calculation method of the API.Slides here
• September, 19th 2022 - Paul Gay, team member.
Title: Deep learning on Graphs
Deep Learning have historically shown success in Speech processing, then tremendous impact on Vision, and eventually on textual data. These models contains many assumptions which are linked to the properties of the data they are applied on (eg. euclidean grid spaces for images). On the other side, many interesting problems can be approach with a graph modelisation, in which these assumptions are no longer valid. In this talk, I wil provide an overview of the last 5 year advancement to apply deep learning principles to learn on graphs.Slides here
• September, 13th 2022 - Florian Valade, Phd FUJISTU-Université Gustave Eiffel.
Title: The rejection option for classification
Prediction confidence is an important issue in Deep Learning. In this talk, I'll cover a theory grounded statistical method, where given a rejection rate, a unlabeled dataset is used to calibrate a probability distribution to perform the rejection/acception of the prediction of any model.
August, 31st 2022 - Jordy Palafox & Matthieu François, team members.
Title: Conference at MAS days in Rouen, Adaptive computing clusters & MCMC algorithms.
August, 1st 2022 - Thomas Poupon, former team member
Title: Presentation of Web App with content on ecology using NLP.
• June, 20th 2022 - Sébastien Loustau & Lorette Duris, team member.
Title: Introduction to the agro-ecological transition and the MSV (Cultivation on living soil, Maraîchage sur Sols Vivants in french)
Agriculture is an essential lever for action in the face of the transformations we are experiencing. The MSV (Maraîchage sur Sols Vivants) is a set of practices of the Agriculture of the living which consists in feeding the soil with organic matter (with high C/N) to support the biodiversity of the soil, and an organic ploughing of the earthworms. After an introduction to MSV, I will describe the status of the 'From Soil to Plate' project, which aims to lead a positive impact project within the university by developing technical and social innovations in favor of these environmentally virtuous practices. Finally, Lorette will explain her work in biochemistry, which consists in analyzing the soils (organic matter rate, ph in particular) and vegetables (dry matter) of different producers from three types of agriculture: conventional, organic and MSV.
• June, 13th 2022 - Paul Gay, team member.
Title: Early Exit YOLO in video object detection
The massive amount of computation required to process video can be largely reduced by discarding redundant information in nearby frames and relying on light AI (quantification, pruning, distillation). In this line, this seminar advocate the potential of Early Exit which enable to dynamically adapt the amount of computation to the difficulty of the detection in each image. On one hand, trivial images can be quickly discarded, and the computation budget can be spent on building good features for hard cases, which will be reused for the next frames. This seminar presents an implementation of these ideas with YOLOv5 on video object detection.Slides here
• May, 23th 2022 - Jordy Palafox, former team member.
Title: Online learning for adaptive learning.
In this talk, we are takling about reduction of the adaptative learning pipeline consumption at Prof En Poche based on unsupervised algorithms. After some theorical reminders on online algorithms, we present a solution using them and illustrate with the first results.
• May, 16th 2022 - Youen Chéné, Founder of Webvert, Web site repair and decarbonization.
Title: State of the art and limits of the measurement of the footprint of a website.
In 2022, there is now multitude ways of measuring and evaluating the environmental impact of a website, such as modeling or measurement tools. They can each evaluate a different perimeter. How to find your way around and what are the current limits?Slides here
• May, 9th 2022 - Matthieu François, team member.
Title: AI for crisis management - Media management part 2, application.
Crisis management requires the processing of a large amount of heterogeneous information in particular audiovisual streams coming from media and social networks. This seminar follows a first part focused on the presentation of open-source models for audio, image, and text analysis. The practical application of these methods will be the basis of this presentation. We will first go back to the motivations for such a process. Then, we will discuss the global architecture adopted with a comparative assessment of methods evaluated on the REPERE annotated data corpus [Giraudel et al]. Finally, we will conclude with a demonstration of a first prototype allowing us to observe and summarize in real-time a news flow.Slides here
• May, 2nd 2022 - Matthieu François, Simon Lebeaud, Nicolas Tirel, team members.
Title: FeedBack about the 5th International School on Deep Learning DeepLearn 2022 Spring.
During the Deep Learn 2022 spring school, Matthieu, Nicolas and Simon had the opportunity to learn more about the most recent advances in the field of deep learning. Aimed at researchers and PhD students in the field of AI, this school presents courses on various fields of application, from weather and health to advanced physics. We also learn more about problems that link several domains such as the robustness and interpretability of a model, or the ability to protect data from malicious attack. For this seminar, we will introduce the school, and then we will each come back on a subject that has marked us, presenting the stakes and the way in which the various challenges were addressed.Link to the website of the conference Slides here
• May, 2nd 2022 - Fatou Kine Sow, team member.
Title: Trash detection and tracking system for recycling awareness in schools
To raise awareness about recycling in schools, prof en poche decided to set up a real-time waste detection system. We will see during this seminar, how to reuse state of the art object detection models such as YOLOV5 and SSDMobilenet V2 to locate and recognize the waste found by the student. We will also see the results with these two models obtained in embedded on a smartphone and in cloud access.
• April, 11th 2022 - Simon Lebeaud, team member.
Title: Computer Vision AI applied to biodiversity and fish tracking in a fish pass context
In the past decade, breakthroughs have been made in object detection due to the massive adoption of Deep convolutional networks. It's time today to use this technology at the service of biodiversity studies. Hizkia has designed and produced a video counting system able to be deployed in any fish pass. We will show how we can now use the data accumulated in the past years to detect and recognize big migratory species to facilitate the counting and study of fish populations throughout France. We will see the potential of state of the art object detection model that we can obtain good results with not a lot of data. You will also have an overview of problematics around river waters imaging.Slides here
• March 28th, 2022 - Nicolas Tirel, team member.
Title: Children Speech Recognition system in a classroom context with energy consumption consideration.
Speech recognition is a complex subject that requires attention to many points during implementation depending on the use case. The choice of architecture, the data collection, their use but also the energy consumed during training are all points addressed during this seminar. Within the project with the startup Prof en Poche, we will see a solution using DeepSpeech which aims to recognize children's voices in the classroom answering mathematical games. This solution comes with a dashboard to display and compare several models via their parameters, data used, energy consumption and especially according to their results. We will see how we were able to obtain up to 18% WER (Word Error Rate) and 12% CER (Character Error Rate) on our target data.Slides here
• March, 21st 2022 - Romain Carrausse, member of TREE lab.
Title: Contestations, conceptions and modes of government of ecological modernization. A reading from a political geography based on the cases of agrivolatism and energy communities.
Romain Carrausse is a post-doctoral fellow at the TEEN research chair and contributes to SoWeSI, a research project conducted in partnership between the company Total Energies and the UPPA. He is working on how ecological modernization is governed and contested at different scales and on different objects. Two works in progress will be presented. The first is on agrivoltaics: a production system that combines energy production, using photovoltaic panels, and agriculture on the same plot of land. We will discuss how the energy sector legitimizes this innovation through a process of internalization of the criticism and demonstration of the agro-economic benefits of the shade produced by the photovoltaic panels. The second is about energy communities. In a context where renewable energy projects face numerous oppositions, locally governed projects, grouped under the term energy community, are experiencing a social and political dynamic in France. A multiscalar approach will be used to analyze the strategies and concepts of energy transition that are reflected in the institutionalization and growth of energy communities.
• March, 14th 2022 - Yanis Chaigneau, former team member.
Title: Machine Learning algorithms for the prediction of the blooming honey plants.
In order to help the beekeepers to plan their transhumances, an accurate forecast of the blooming of the honey plants is required. This seminar focuses on the use of machine learning algorithms for phenology. An innovative non-supervised algorithm is presented and compared to classical approaches.
• February, 14th 2022 - Sébastien Lousteau, team member.
Title: MCMC optimization for high-dimensional problems.
The aim of this talk is to introduce a greedy MCMC optimizer for Deep Learning. After a gentle start about the convergence of standard Metropolis Hasting algorithm, and discussion about MCMC alternatives, I will present recent MCMC challenging algorithms for recent high dimensional machine learning problems, where the dimension of the Markov Chain could change over time. Then, I will describe how to adapt these ideas to build a new optimizer for Deep Learning and shows its nice properties to learn sparse deep nets, as well as the next challenges to have a competitive counterpart of standard stochastic gradient methods.
• January, 31st 2022 - Paul Gay, team member.
Title: Social computing for environnment applications.
This seminar is an introduction to social computing with a focus on environnemental applications ie the behavior or the resilience of a socio-economic object when triggered by an environnemental hazard. This subject encompasses for example studying a crowd opinion to a new policy through social network analysis, or broadcast media representation thanks to NLP and graph analysis methods. After a general introduction, the talk will focus on two case studies : The artificial intelligence on crisis management, and the community based energy storage planning to improve photovoltaic system adoption.Slides here
• January, 17th 2022 - Jordy Palafox, former team member.
Title: Clustering and recommandation system for adaptative learning
In this talk, we give some details about code optimization of the MathIA engine realized by Prof en Poche. We explore some clues about k-medoids clustering methods. After that, we take care about the consumption of the algorithm measured by AIPowerMeter compared to the real consumption obtained with Schneider Power meters.
January 2022, 3rd 2022, Seminar of Sébastien Loustau, and Team progress summary with new team members !
• December, 20th 2021 - Matthieu François, team member.
Title: AI in crisis management - Media management with AI.
Crisis management requires the management of a large amount of heterogeneous information. This seminar focuses on the processing of media (TV news, tweets), whose monitoring is crucial for decision makers to understand the perception of a crisis by the general public. It will be presented first efforts to build a monitoring system: (i) open-source libraries are used to extract speech and text embedded in videos (ii) Named-Entity Recognition is applied to bring a summary of the information to a human operator. In particular, an important issue is the construction of a model able to detect concepts appearing during the crisis and thus unknown to the model at the time of learning. We will study three complementary strategies based on regular expressions, word embedding comparisons and a BERT model learned online as training data arrives.Slides here
December, 6th 2021 - Simon Gosset & Martin Chauvin, CarbonAI team.
Title: CarbonAI an opensource project presentation.
November, 29th 2021 Seminar for the Approximate Bayesian Inference Team
Title: Deep learning theory for power-efficient algorithms.
• November, 22th 2021 - Sébastien Loustau, team member.
Title: Forget SGD - Deep learning theory for a new optimizer
In this talk, I will introduce alternatives to standard gradient descents to learn power-efficient deep learning algorithms. After a gentle start about optimization with mirror descents, we present recent theoretical advances on Pac-Bayesian sparse deep learning, leading to a new greedy optimizer to learn sparse and potentially binarized deep networks. We finally introduce new divergences to the prior, and rely this framework with metric task systems.
November, 17th 2021 Workshop at ACML, Power-Efficient Deep Learning.
Organized with Pierre Alquier from the Approximate Bayesian Inference Team at Riken Institute.
• November, 8th 2021 - Jordy Palafox, former team member.
Adaptive Learning and Clustering methods
In the context of adaptive learning, clustering methods are used to recognize students with the same profil. Here, we will focus on the clustering algorithm used by Prof en Poche which is a combinaison of two methods : the KMedoids and the Louvain algorithm to obtain a robust method. After introducing it, we will measure the consumption of the algorithm and explore how to reduce it. We will conclude with some recent methods using deep learning based on autoencoders.
• October, 25th 2021 - Paul Gay, team member.
Title: Power efficient transformers
As transformers are becoming the standard NLP tool, questions are raised about ethics, bias and energy consumption. This last topic is of importance as these models are the biggest ones in the large and hungry power deep learning model trend. In this seminar, I will present in the first part a conmprehensive tutorial on the principles of attention and how the transformers have conquered the state of the art. This details will equip us for the second part in which I will survey a number of methods which are concerned in making the transformers lighter and more available for practicionners with low computation resources.Slides here
• October, 11th 2021 - Matthieu François, team member.
Title: AI addition to crisis management
Human impact on our planet is increasing the scale and the number of environmental disasters. During this seminar I'll present our join project with Altanoveo. This project is about AI integration to climate or industrial crisis management methods. I will describe the potential of IA in this domain and present two models on tweet classification and fire detection on natural images.Slide here
• September, 27th 2021 - Sébastien Loustau, team member.
Title: Introduction to convex optimization
In this lecture, I will introduce convex optimization theory and mirror descent. We start with a theoretical motivation and studyt of (stochastic) gradient descent, and then moove to the non-euclidean setting to derive mirror descent algorithm as a generalization of gradient descent. We finally apply it to the context of expert advices to recover the classical regret bound for exponential weighted averages previously presented in the first seminar in july, and discuss possible applications to Green AI.Slide here
• September, 13th 2021 - Sébastien Loustau, team member.
Title: Kick Off GreenAI Uppa
Official kick off of the GreenAI UPPA project ! After presenting the climatic and mathematical motivations (has the earth ever been this hot before ?), we introduce the context and support we have from the public institutions and the SMEs. We explain how the team will be organized, and inspired from the best of both worlds. Then we take 30 minuts to write our elevator pitch. Welcome to Jordy and Matthieu !Slide here
• August, 23rd 2021 - Paul Gay, team member.
Title: Measuring the Power draw of computers
The ability of measuring power and consumption of machine learning algorithms is necessary to design new efficient ones. Nowadays, there is a growing interest in the machine learning and IT community for measuring the consumption at different scale, from the AI model to the entire data center. In this talk, we survey recent tools and softwares based on RAPL and NVIDIA-SMI and highlight the dependancy to the hardware considered (CPU, GPU), as well as the different sources of consumption in the architecture of a computer. The final goal is to give to engineers and data scientists the capacity to measure the consumption of deep/machine learning algorithms via our open source software deep_learning_power_measure developed by Green AI Uppa.Slide here
• July, 26th 2021 - Julien Mercier, team member.
Title: How to binarize a neural network
In this talk, I propose to present the main pros and cons of binarization via the gradient. We present two main historical attempt: BinaryConnect and BinaryNetwork, based on two recent papers.Slide here
• July, 5th 2021 - Sébastien Loustau, team member.
Title: How to penalize deep learning with power measurements ?
In this talk, I propose to introduce the main theoretical foundations of online learning and PAC-Bayesian theory and how it could be used to build new power-efficient algorithms. After a gentle start dedicated to the problem of prediction with expert's advices, I will present the PAC-Bayesian paradigm and how it is related to the context of aggregation of expert's advices and stochastic algorithms. We apply this theory to learn sparse deep nets where the networks are coming from recent advances in binarization (BinaryConnect, XNor-nets, Xnor-nets ++). We finally sketch how to generalize these results to more suitable divergences such as Optimal Transport, a nice and promising field in order to measure the cost of choosing sequentially a particular algorithm in terms of electric consumption. This talk is based on the two following papers: Learning with BOT and Sparsity regret bounds for XNOR-Nets.Slide here