We present an overview of recent advances in the understanding of optical beams in nonlinear media with a spatially nonlocal nonlinear response. We discuss the impact of nonlocality on the modulational instability of plane waves, the collapse of finite-size beams, and the formation and interaction of spatial solitons.

We show that quadratic solitons are equivalent to solitons of a nonlocal Kerr medium. This provides new physical insight into the properties of quadratic solitons, often believed to be equivalent to solitons of an effective saturable Kerr medium. The nonlocal analogy also allows for novel analytical solutions and the prediction of novel bound states of quadratic solitons.

We study the formation and interaction of spatial dark optical solitons in materials with a nonlocal nonlinear response. We show that unlike in local materials, where dark solitons typically repel, the nonlocal nonlinearity leads to a long-range attraction and formation of stable bound states of dark solitons.

We numerically study supercontinuum (SC) generation in photonic crystal fibers pumped with low-power 30-ps pulses close to the zero dispersion wavelength 647nm. We show how the efficiency is significantly improved by designing the dispersion to allow widely separated spectral lines generated by degenerate four-wave-mixing (FWM) directly from the pump to broaden and merge. By proper modification of the dispersion profile the generation of additional FWM Stokes and anti-Stokes lines results in efficient generation of an 800nm wide SC. Simulations show that the predicted efficient SC generation is more robust and can survive fiber imperfections modelled as random fluctuations of the dispersion coefficients along the fiber length.

We explore the suitability of self-attention models for character-level neural machine translation. We test the standard transformer model, as well as a novel variant in which the encoder block combines information from nearby characters using convolutions. We perform extensive experiments on WMT and UN datasets, testing both bilingual and multilingual translation to English using up to three input languages (French, Spanish, and Chinese). Our transformer variant consistently outperforms the standard transformer at the character-level and converges faster while learning more robust character-level alignments. 1