More Like this

1. Leveraging Domain Information for the Efficient Automated Design of Deep Learning Accelerators

   https://doi.org/10.1109/HPCA56546.2023.10071095  

   Sakhuja, Chirag; Shi, Zhan; Lin, Calvin (February 2023, International Symposium on High Performance Computer Architecture)

   Deep learning accelerators are important tools for feeding the growing demand for deep learning applications. The automated design of such accelerators--which is important for reducing development costs--can be viewed as a search over a vast and complex design space that consists of all possible accelerators and all the possible software that could run on them. Unfortunately, this search is complicated by the existence of many ordinal and categorical values, which are critical to explore for the ultimate design but are not handled well by existing search techniques. This paper presents a technique for efficiently searching this space by injecting domain information--in this case information about hardware/software (HW/SW) co-design--into the automated search process. Specifically, this paper introduces a novel Bayesian optimization framework called daBO (domain-aware BO) that accepts domain information as input, including those describing ordinal and categorical values. This paper also introduces Spotlight, a design tool based on daBO, and this paper empirically shows that Spotlight produces accelerator designs and software schedules that are orders of magnitude better than those created by the state-of-the-art. For example, for the ResNet-50 deep learning model, Spotlight produces a HW/SW configuration that reduces delay by 135x over the configuration produced by ConfuciuX, a state-of-the-art HW/SW co-design tool, and Spotlight reduces energy-delay product (EDP) by 44x over an Eyeriss-like accelerator, which is an edge-scale hand-designed accelerator. In the realm of cloud-scale accelerators, Spotlight reduces the EDP of a scaled-up Eyeriss-like accelerator by 23x. Our evaluation shows that Spotlight benefits from the efficiency of daBO, which allows Spotlight to identify accelerator designs and software schedules that prior work cannot identify. 

   more » « less
2. Interactive Learning from Policy-Dependent Human Feedback

   MacGlashan, James; K Ho, Mark; Loftin, Robert; Peng, Bei; Wang, Guan; Roberts, David L.; Taylor, Matthew E.; Littman, Michael L. (July 2017, ICML)

   This paper investigates the problem of interactively learning behaviors communicated by a human teacher using positive and negative feedback. Much previous work on this problem has made the assumption that people provide feedback for decisions that is dependent on the behavior they are teaching and is independent from the learner’s current policy. We present empirical results that show this assumption to be false—whether human trainers give a positive or negative feedback for a decision is influenced by the learner’s current policy. Based on this insight, we introduce Convergent Actor-Critic by Humans (COACH), an algorithm for learning from policy-dependent feedback that converges to a local optimum. Finally, we demonstrate that COACH can successfully learn multiple behaviors on a physical robot. 

   more » « less
3. Incoherence and fibering of many free-by-free groups

   https://doi.org/10.5802/aif.3494  

   Kropholler, Robert P; Walsh, Genevieve S (January 2022, Annales de l'Institut Fourier)

   A group is called free-by-free if it is the semi-direct product of two finitely generated free groups. A group is coherent if any finitely generated subgroup is finitely presented, and incoherent otherwise. In this paper, the authors provide evidence towards the conjecture (due independently to the authors and Dani Wise) that every free-by-free group is incoherent. To do this, they give a homological condition which lets them conclude that the free-by-free group has a finite index subgroup which surjects onto ℤ with finitely generated kernel; standard arguments imply that this kernel cannot be finitely presented. As an important special case, they show that if the free-by-free group is hyperbolic and virtually special, then it is incoherent. 

   more » « less
4. The traveling salesman theorem for Jordan curves

   Bishop, Christopher J. (September 2023, Advances in mathematics)

   This gives an improvment of Peter Jones's traveling salesman theorem that holds for Jordan curves, but not for general sets. His theorem implies that the length of a Jordan arc is bounded by (1+delta)diameter + C(delta) beta-sum, and this paper shows this can be replaced by chord + O(beta-sum), where O(diameter) is replaced by the distance between the endpoints of the curve. This is true in all finite dimensions (with a dimension dependent constant). A corollary of our self-contained argument proves the ususal TST in all dimensions (a result of Okikiolu). An appendix proves a folklore result that several different formulation of Jones's theorem are all equivalent. 

   more » « less
5. Reversible Electronic Patterning of a Dynamically Responsive Hydrogel Medium

   https://doi.org/10.1002/adfm.202302549  

   Yang, Chen; Liu, Yi; Wang, Manya; Hu, Hui; Zhao, Zhongtao; Deng, Hongbing; Payne, Gregory_F; Shi, Xiaowen (May 2023, Advanced Functional Materials)

   Abstract A dynamically responsive hydrogel medium is prepared from two self‐assembling components, a polysaccharide (chitosan) and a surfactant (sodium dodecyl sulfate; SDS). It is shown that this medium can be patterned using an electrode “pen” to reconfigure supramolecular structure: cathodic writing induces neutral chitosan chains to form a crystalline network, while anodic writing generates cationic chitosan chains that electrostatically crosslink with anionic SDS micelles. Both supramolecular structures are re‐configurable and each is stabilized by structure‐induced shifts in chitosan's pKa, thus electronically written patterns can be erased, new patterns can be written, and patterns can be written in three dimensions. Further, it is shown that NaCl‐induced morphological transitions of the SDS micelles allow patterns to be reversibly concealed or revealed. To demonstrate the versatility of this medium for information storage, a quick response (QR) code is electronically written and it is shown that this code can be recognized by a standard cellphone app. This QR code can be concealed by making the medium opaque (i.e., by obscuring the pattern) or by making the pattern evanescent (i.e., by making pattern invisible). Overall, this work demonstrates that a dynamically responsive medium composed of simple, safe and sustainable components can be reversibly patterned with spatial and quantitative control using top‐down electronic inputs. 

   more » « less