AI’s Smarts Now Come With a Big Price Tag
Will Knight | Wired Magazine
As language models get more complex, they also get more expensive to create and run. One option is a startup, Mosaic ML, spun out of MIT that is developing software tricks designed to increase the efficiency of machine-learning training.
PointAcc: Efficient Point Cloud Accelerator
Tuesday, October 19, 2021 | 3:00pm – 4:15pm ET
Speakers: Yujun Lin and Song Han, MIT
Natural Language Processing Accelerator for Transformer Models
Song Han, Anantha Chandrakasan
This project aims to develop efficient processors for natural language processing directly on an edge device to ensure privacy, low latency and extended battery life. The goal is to accelerate the entire transformer model (as opposed to just the attention mechanism) to reduce data movement across layers.
Boltzmann Network with Stochastic Magnetic Tunnel Junctions
Luqiao Liu, Marc Baldo
Networks formed by devices with intrinsic stochastic switching properties can be used to build Boltzmann machine, which has great efficiencies compared with traditional von Neumann architecture for cognitive computing due to the benefit from statistical mechanics of building blocks.
Electrochemistry and Material Science of Proton-based Electrochemical Synapses
Bilge Yildiz, Ju Li
Electrochemical ionic-electronic devices have an immense potential to enable a new domain of programmable hardware for machine intelligence.
A Framework to Evaluate Energy Efficiency and Performance of Analog Neural Networks
Vivienne Sze, Joel Emer
This project pursues an integrated framework that includes energy-modeling and performance evaluation tools to systematically explore and estimate the energy-efficiency and performance of Analog Neural Network architectures with full consideration of the electrical characteristics of the synaptic elements and interface circuits.
CMOS-Compatible Ferroelectric Synapse Technology for Analog Neural Networks
Jesús del Alamo
This research project investigates a new ferroelectric synapse technology based on metal oxides that is designed to be fully back-end CMOS compatible and promises operation with great energy efficiency.
Crossing the Hardware-Software Divide for Faster AI
Thursday, April 29, 2021 | 12pm - 1pm ET
Panel Discussion: Vivienne Sze, Song Han, Aude Oliva
Using Artificial Intelligence to Generate 3D Holograms in Real-time
Daniel Ackerman | MIT News Office
Despite years of hype, virtual reality headsets have yet to topple TV or computer screens as the go-to devices for video viewing.
Shrinking Massive Neural Networks Used to Model Language
Daniel Ackerman | MIT News Office
Researcher Jonathan Frankle and his “lottery ticket hypothesis” posits that, hidden within massive neural networks, leaner subnetworks can complete the same task more efficiently.