As another example, here the state is similar to the prototypes for "Accelerate" and "Turn Right", so the agent accelerates and turns right. Our human study shows this research has the potential to help users predict out of distribution failures.

The work forces an agent to make decisions based on the current state's similarity to pre-defined human-interpretable prototypes. For example, here the state is similar to the prototypes for "Brake" and "Turn Left", so the agent brakes and turns left.

** New ICLR Spotlight Paper 2023 ** Excited to announce our work building inherently interpretable Deep RL agents that doesn't sacrifice performance, and calibrates appropriate user trust. (details below...) openreview.net/forum?id=hWwY… By @EoinKNNy, @MycalTucker, and @julie_a_shah

Excited to share our latest paper to appear at NAACL this year, ExSum: From Local Explanations to Model Understanding, in collaboration with Marco Tulio Ribeiro from Microsoft Research.

Enjoy a free virtual screening of #PictureAScientist, with a followup panel discussion!

Talk streaming now by Oriol Vinyals (@OriolVinyalsML), head of the deep learning team at @DeepMind: youtube.com/watch?v=kYAd5cV1…

Join IRG at the AAAI StarAI workshop Friday 7th 3-5pm to hear about our efforts to sample prediction-matching examples in neural networks. 1 or 7? 8 or 9? The model isn't sure; sometimes we aren't either. arxiv.org/pdf/2001.03076.pdf @SerenaLBooth, @ankitjs, and Yilun Zhou

The view from our lab yesterday. Fantastic undergrad research presentations in the Stata center.

Dance like everybody's watching! Robotina busts a move to celebrate IRG's new Optitrack system. Human-robot dancing, taught by @jayscheib, coming to a lab near you.

IRG at Amazon Robotics! A bunch of us visited @AmazonRobotics in North Reading to get a better sense of the challenges faced by engineers and associates in mixed human-robot warehouses.