LeCun was born in France in 1960 and has been fascinated with AI since an early age. At the age of 9, he saw A Space Odyssey and that made him fond of space travel, AI and human intelligence. He learnt at young age that intelligence is self-organizing — complex behavior emerges from the interactions of simple elements.
LeCun started his professional work in 1980s. Neural networks then had fallen out of favour. There were limitations of ‘perceptrons’, some of the early neural networks, first pioneered in 1950s.
The field of AI shifted to symbolic and rule-based systems. The field was revived by people not obsessed by history. They established a connection between statical physics, theoretical neuroscience and neural nets.
He did his PhD work in Pierre University in mid-1980s.
He developed the renowned backpropagation algorithm — his first contribution to neural networks. Here the network learns based upon errors detected in its output. These are ‘backpropagated’ through the network to adjust internals weights. It improves accuracy. It has now become an established technique to train the network. He completed his doctorate in 1987.
He joined the Univrsity of Toronto for post-doctoral fellowship under Geoffrey Hinton. A year later he joined Bell Labs. He contributed to the development of CNN — convolutional neural networks. CNNs scan images. They detect features like edges, textures and shapes. They detect these irrespective of where they are in the visual field. It improves computer vision (CV). It revolutionized handwriting recognition, cheque clearing and facial recognition. It helped medical imaging, autonomous vehicle perception and AR.
He also had stints at AT&T and NEC. LeCun later joined New York Univrsity in 2003. He still serves as Silver Professor. He was recruited by Zuckerberg for Facebook in 2013 and is currently Chief AI Scientist.
LLMs have limitations. They are just token generators and do compute of fixed amount to generate a token. It is a reactive system. It is an intuitive system. The slower, deliberate reasoning system of human brain is another system. It is easier to teach AI systems higher order skills, say chess playing and clearing tests — cortical based reasoning skills come much later. These require more cognitive effort. This is called Moravec’s paradox. The machines act exactly in reverse direction. LLMs master higher order skills, say NLP. But they lack foundational abilities. The real world is messy and continuous. Humans process more data than AI systems.
LeCun is pioneering an alternative approach Joint Embedding Predictive Architecture (JEPA) that mimic that physical world on visual output. The predictions do not happen in the space of raw sensory inputs. They happen in abstract representational space. In another V-JEPA or video JEPA model, LeCun’s team trained a system to complete partially occluded videos.
LLMs are probabilistic while predicting next token. LeCun’s system learn to represent the world at multiple levels of abstraction and predicts how these representations evolve under different conditions.