Paper on egocentric zone-aware action recognition across environments just published in Elsevier Patter Recognition Letters
In our daily life, we tend to associate tools and objects with the actions they can facilitate (like a glass with pouring).
Interestingly, we as humans can assign affordances not only to objects, but also to places. For example, we typically wash
things near a sink, rather than by the kitchen stove. This prior knowledge allows us to use appliances in a friend’s house,
without much thought.In our new paper, we demonstrate how deep learning models for egocentric action recognition can exploit conceptual
activity-centric zones and use them to improve robustness across domains.
Links:
Project pagePaperIs hashtag#AI a true hashtag#revolution, or just hype of the moment?Well, according to forecasts, AI can have an impact on our lives similar to the industrial revolution.But at what cost? In 2019 researchers estimated that training a (fairly small, with 217M parameters) transformer burns 201 kWh of energy which - in US - results on approx 90 KG of CO2 equivalent emitted.
hashtag#GPT3.5 (the engine of the famous hashtag#chatgpt) has 175 billions of parameters, and its training used 1287 MWh of electricity (approx. 500 tons of CO2e), equivalent to driving 112 gasoline powered cars for a year! [1]How much is that? In 2021 this represented the impact of 33 US citizens’ daily life for a whole year (15 tons each), or 83 Italians (6 tons each) [2] (yes we have a much smaller carbon footprint 😉). Using those models day-by-day may be even more expensive, as 3/5 of al the energy used by machine learning is during inference [3]. Using GPT3.5 produces 8.5 tons of CO2/year [3], much more than a person living in Turin, Naples or Rome.To answer our first question, then, I would say: yes, it is… but our planet cannot afford it!
Building larger and larger models, trained on massive amount of data, is simply not the way to go. We are doing research in Politecnico di Torino to demonstrate that another way is possible. hashtag#Small, hashtag#efficient, hashtag#green and hashtag#custom models are more than enough to solve most of the problems out there. And now, with the help of our newborn Focoos AI we will bring this “gentle” green revolution to the AI market.Stay tuned for a new generation of intelligent, efficient and green machines powered by hashtag#FocoosAI technology. Find us at https://www.focoos.ai/hashtag#GreenAI hashtag#FocoosAI hashtag#PoliTo
I3P - Innovative Companies Incubator of Politecnico Torino[1] https://lnkd.in/dvi9ZJJF
[3] https://lnkd.in/d8tbFZrm
[2] https://lnkd.in/dthCJCcA
[4] https://lnkd.in/d85VBFmi
