Flat electronic bands, characteristic of magic-angle twisted bilayer graphene (TBG), host a wealth of correlated phenomena. Early theoretical considerations suggested that, at the magic angle, the Dirac velocity vanishes and the entire width of the m
oire bands becomes extremely narrow. Yet, this scenario contradicts experimental studies that reveal a finite Dirac velocity as well as bandwidths significantly larger than predicted. Here we use spatially resolved spectroscopy in finite and zero magnetic fields to examine the electronic structure of moire bands and their intricate connection to correlated phases. By following the relative shifts of Landau levels in finite fields, we detect filling-dependent band flattening, that unexpectedly starts already at ~1.3 degrees, well above the magic angle and hence nominally in the weakly correlated regime. We further show that, as the twist angle is reduced, the moire bands become maximally flat at progressively lower doping levels. Surprisingly, when the twist angles reach values for which the maximal flattening occurs at approximate filling of $-2$, $+1$,$+2$,$+3$ electrons per moire unit cell, the corresponding zero-field correlated phases start to emerge. Our observations are corroborated by calculations that incorporate an interplay between the Coulomb charging energy and exchange interactions; together these effects produce band flattening and hence a significant density-of-states enhancement that facilitates the observed symmetry-breaking cascade transitions. Besides emerging phases pinned to integer fillings, we also experimentally identify a series of pronounced correlation-driven band deformations and soft gaps in a wider doping range around $pm 2$ filling where superconductivity is expected. Our results highlight the role of interaction-driven band-flattening in forming robust correlated phases in TBG.
Magic-angle twisted bilayer graphene (MATBG) exhibits a range of correlated phenomena that originate from strong electron-electron interactions. These interactions make the Fermi surface highly susceptible to reconstruction when $ pm 1, pm 2, pm 3$ e
lectrons occupy each moir e unit cell and lead to the formation of correlated insulating, superconducting and ferromagnetic phases. While some phases have been shown to carry a non-zero Chern number, the local microscopic properties and topological character of many other phases remain elusive. Here we introduce a set of novel techniques hinging on scanning tunneling microscopy (STM) to map out topological phases in MATBG that emerge in finite magnetic field. By following the evolution of the local density of states (LDOS) at the Fermi level with electrostatic doping and magnetic field, we visualize a local Landau fan diagram that enables us to directly assign Chern numbers to all observed phases. We uncover the existence of six topological phases emanating from integer fillings in finite fields and whose origin relates to a cascade of symmetry-breaking transitions driven by correlations. The spatially resolved and electron-density-tuned LDOS maps further reveal that these topological phases can form only in a small range of twist angles around the magic-angle value. Both the microscopic origin and extreme sensitivity to twist angle differentiate these topological phases from the Landau levels observed near charge neutrality. Moreover, we observe that even the charge-neutrality Landau spectrum taken at low fields is considerably modified by interactions and exhibits an unexpected splitting between zero Landau levels that can be as large as ${sim },3-5$ meV. Our results show how strong electronic interactions affect the band structure of MATBG and lead to the formation of correlation-enabled topological phases.
This paper analyzes the effects of approximate multiplication when performing inferences on deep convolutional neural networks (CNNs). The approximate multiplication can reduce the cost of the underlying circuits so that CNN inferences can be perform
ed more efficiently in hardware accelerators. The study identifies the critical factors in the convolution, fully-connected, and batch normalization layers that allow more accurate CNN predictions despite the errors from approximate multiplication. The same factors also provide an arithmetic explanation of why bfloat16 multiplication performs well on CNNs. The experiments are performed with recognized network architectures to show that the approximate multipliers can produce predictions that are nearly as accurate as the FP32 references, without additional training. For example, the ResNet and Inception-v4 models with Mitch-$w$6 multiplication produces Top-5 errors that are within 0.2% compared to the FP32 references. A brief cost comparison of Mitch-$w$6 against bfloat16 is presented, where a MAC operation saves up to 80% of energy compared to the bfloat16 arithmetic. The most far-reaching contribution of this paper is the analytical justification that multiplications can be approximated while additions need to be exact in CNN MAC operations.
