tinyML Talks on July 1, 2021 “State of the TinyML today (Applications, état de l'art et enjeux du TinyML)” by Frédéric Pétrot, Etienne Balit and Loic Lietar

tinyML Talks
1

We held our next tinyML Talks webcast. Frédéric Pétrot, Etienne Balit and Loic Lietar presented State of the TinyML today (Applications, état de l’art et enjeux du TinyML) on July 1, 2021.

Banner
Banner1280×220 58.4 KB

The 1st TinyML meetup will be held virtually as an open panel discussion with the keynote speakers. We would like to discuss the landscape and potential of today’s ultra-low-power applications, recent advances and challenges in TinyML. We will encourage questions, opinions and suggestions from the audience on this subject.

Frédéric Pétrot est docteur en informatiquede l’Université Pierre et Marie, Paris, France, depuis 1994. Il arejoint Grenoble INP/Ensimag en 2004 en tant que Professeur. Ses travaux portentsur la conception et l’implantation des systèmes numériques intégrés, àusage général ou spécialisés, par exemple pour l’accélération del’intelligence artificielle.

Etienne Balit est Directeur R&D à Neovision. Issu de l’Ecole NormaleSupérieure où il a obtenu un master en modélisation pour les sciencescognitives, il a ensuite réalisé une thèse en informatique à Inria, où il adéveloppé des connaissances pointues en intelligence artificielle etparticulièrement en réseaux de neurones. A Neovision, il coordonne l’équipeR&D qui réalise une veille constante des dernières avancées scientifiqueset techniques et qui investit sur des sujets porteurs, afin d’apporter auxentreprises des innovations performantes et à l’état de l’art.

Loïc is a co-founder and the CEO of GreenWaves Technologies. GreenWaves Technologies, a fabless semiconductor company, designs disruptive ultra-low power embedded solutions for interpreting and transforming rich data sources such as images, sounds, radar signatures and vibrations using AI and signal processing in highly power-constrained devices such as hearables, wearables and IoT sensors. Prior to this, Loic worked 25 years for ST where he led several product divisions, has been the Chief Strategy Officer and co-founded and managed ST’s corporate venture fund. Loïc has been an active business angel for the last 8 years. He has also been president of Minalogic, the French cluster for semiconductors.

==========================

Watch on YouTube:
Frédéric Pétrot, Etienne Balit and Loic Lietar

Download presentation slides:
Frédéric Pétrot, Etienne Balit and Loic Lietar

Feel free to ask your questions on this thread and keep the conversation going!

3

Is there a point at which increasing layers becomes less efficient than increasing bytes? Do you have a curve?
The team of Michaela Blott at Xilinx has published a few papers in which you can find some curves. Note that this is really dependent on the NN architecture, so there is no general rule as far as I know.

Coming back on distributed AI topic, but more than “simple wake up mechanisms”, are there any advanced research on that taking into account energy budget for communication?
Well, there are lots of papers related to that, and Sigfox and LoRa have been precisely designed to mitigate communication consts. I can suggest having a look at this paper for an overview : http://mespages.univ-brest.fr/~bounceur/anr/persepteur/articles/atsip_2017_noreen.pdf

The trade-off zone seems not to include any safety-critical applications. Is this intentional? Should we consider safety-critical applications as impossible tasks or do you forsee any tasks in this domain to be possible to some extent in the near future?
ML in general remains a bit of black magic, which already doesn’t go well with safety-critical applications. So even more in the context of TinyML where trade-offs are about cost vs accuracy. It will take time :slight_smile:

Does the change going on in new processors mean tinyML programmers will be doing a lot of education vs. work. What percent over the next 5 years?
The trend is for SW tools to enable the programmer to harness the efficiency of a given processor without having to master the sophistication of its architecture. So specific education shouldn’t be required, compared to other embebbed programming. Education is and will be about mastering data production and model selection to achieve a given project goal. And as long as the state-of-the-art will be rapidly evolving as it is the case today (probably still for few years) remaining closely informed of its progress will be necessary. I’d say 15% of one time, if a number has to be given.

How do you solve RAM limitations on microcontrollers? If you have just 256kb on it?
For some applications, small NNs, well pruned, can be quite efficient. You might take a look at this paper :

It seems many companies that embark on the design of a product with embedded AI underestimate the investment to correctly train the models they’ll use ** taking into account their own specific use case** (esp. corner cases).
Would you agree, based on your experience ?
And if so :

  • do you think this may discourage some of them after a challenging or ill-fated experience ?
  • what avenues could be envisaged to decrease the required investment ?
    Advanced technology needs knowledable people until tools are ready, and we are not there yet. How to choose a NN for a given application is more magic than science today, and, for example, companies are resorting to “network architecture search” by trying zillions of possibilities and hoping one will pop out well. At the cost of MW of power consuption, though, not really TinyML.

Would there be a tiny-ASIC HW acclerator on or added to MCU to enable better performance?
There are already CNN accelerators in some tinyml mcu on the market (see maxim integrated or GreenWaves for example).