Introduction to Energy-Based Learning | Yann LeCun Paper

Key Takeaways

one of the biggest problem with artificial intelligence is that it's missing the ability to predict what's going to happen in the world if you train a system to drive a car you don't want it to do stupid things like falling off a cliff over and over again learning slowly by trial and error that it's not a good idea this deep learning model may be a way to make machines learn more like us [Music] when training a machine learning model you want to predict what's going to happen in the future so it won't try stupid things and will improve more easily in a human way most of the success of deep learning to date have been due to both supervised and reinforcement learning but it requires too many trials to be applicable in the real world for example if you train a system to drive a car you don't want it to drive into a tree or fall off a cliff over and over again learning slowly by try on error that it's not a good idea if you are not convinced yet the program alpha star took around 200 years of equivalent real-time play to defeat the starcraft pro players using reinforcement learning and this was achieved on a single map in a single type of players you can see that this is clearly not viable for real world applications since you can't run the real world faster than the real time if you want our machines to have a human level intelligence we first need to understand how the humans and animals learn this is why energy based learning along with self-supervised learning which i covered in a previous video that i linked in the description has been popularized by yan likum we want our machines to learn like us using a few or more the right number of examples and predicting what may happen in the future just like babies do using this technique you can predict what's going to happen in the world for example in videos by masking a part of the video and ask the system to reconstruct the next few frames it is applicable in natural language processing too where you mask the end of the phrase and let the system try to reconstruct it but there is no way to make a system predict exactly what's going to happen so we have to find a way to deal with the unsureness of the prediction results but now you may wonder how can we achieve this energy based learning this is done by learning an energy function that takes low values on the data manifold and higher values everywhere else in this example a model measures the compatibility between observed variables x and variables to be predicted y using an energy function the model could be viewed as an energy function measuring the goodness of each possible configuration of x and y in short the model asks itself what is the y that is the most compatible with this x so the model autonomously pushes down on the energy of the desired outputs and pushes up on everything else orienting the model to a low energy output where the future guesses are more plausible you can achieve that by using two different objective functions with a gun architecture which i covered in my last video and is linked in the description then you need to build the machine so that the volume of the low energy stuff is constant using an algorithm like king means pca or gmm more information about these algorithms are linked in the description as well finally auto encoders are used to train a dynamic system and a regularizer that limits the volume of space that has low energy limiting the possibilities of the output guesses this new type of learning has great potential and can be already used in different fields of machine learning from self-driving car applications reconstructing some part of a video predicting the words missing in a damaged boot face recognition image segmentation and more this was just an introduction to learn more about this check out the links in the description please leave a like if you learned something and subscribe to the channel to not miss any further terms clearly explained [Music] you