The problem of cross-modal retrieval, e.g., using a text query to search for images and vice-versa, is considered in this paper. A novel model involving correspondence autoencoder (Corr-AE) is proposed here for solving this problem. The model is constructed by correlating hidden representations of two uni-modal autoencoders. A novel optimal objective, which minimizes a linear combination of representation learning errors for each modality and correlation learning error between hidden representations of two modalities, is used to train the model as a whole. Minimization of correlation learning error forces the model to learn hidden representations with only common information in different modalities, while minimization of representation learning error makes hidden representations are good enough to reconstruct input of each modality. A parameter α is used to balance the representation learning error and the correlation learning error. Based on two different multi-modal autoencoders, Corr-AE is extended to other two correspondence models, here we called Corr-Cross-AE and Corr-Full-AE. The proposed models are evaluated on three publicly available data sets from real scenes. We demonstrate that the three correspondence autoencoders perform significantly better than three canonical correlation analysis based models and two popular multi-modal deep models on cross-modal retrieval tasks.

The ICML 2013 Workshop on Challenges in Representation Learning(1) focused on three challenges: the black box learning challenge, the facial expression recognition challenge, and the multimodal learning challenge. We describe the datasets created for these challenges and summarize the results of the competitions. We provide suggestions for organizers of future challenges and some comments on what kind of knowledge can be gained from machine learning competitions.

Abstract. The ICML 2013 Workshop on Challenges in Representation Learning3 focused on three challenges: the black box learning challenge, the facial expression recognition challenge, and the multimodal learning challenge. We describe the datasets created for these challenges and summarize the results of the competitions. We provide suggestions for organizers of future challenges and some comments on what kind of knowledge can be gained from machine learning competitions.

We investigated the oxidation of CH4 on oxygen-pre-covered IrO2(110) surfaces using temperature-programmed reaction spectroscopy (TPRS) and density functional theory (DFT). Our TPRS results show that on-top oxygen (Oot) species hinder CH4 adsorption, providing evidence that CH4 adsorbs on coordinatively unsaturated Ir atoms. We also find that the fractional yield of adsorbed CH4 that reacts during TPRS remains constant at ∼70% as the Oot-coverage increases to about 0.5 monolayer for saturation CH4 coverage, demonstrating that O-rich IrO2(110) surfaces are highly active in promoting CH4 C–H bond cleavage. Our results show that Oot atoms promote CH4 oxidation to CO2 as well as H2O formation while suppressing CO and recombinative CH4 desorption, as evidenced by an increase in the fractional yield of CO2 produced during TPRS and a downshift of CO2 and H2O TPRS peak maxima with increasing Oot-coverage. DFT predicts that initial CH4 bond cleavage is highly facile on both stoichiometric and O-rich IrO2(110) and can occur by either H-transfer to an Oot or a bridging O-atom of the surface. Our calculations also predict that oxidation of the CH x species that result from CH4 activation is more facile on O-rich compared with stoichiometric IrO2(110), and that complete oxidation is strongly favored on the O-rich surface, in good agreement with our experimental findings. According to the calculations, key steps in the CH4 oxidation pathway have significantly lower-energy barriers on O-rich vs stoichiometric IrO2(110) because these steps involve reaction with Oot atoms initially present on the surface rather than the abstraction of more strongly bound Obr species. High coverages of O-atoms also enable adsorbed intermediates to oxidize extensively on O-rich IrO2(110), without the intermediates needing to overcome diffusion barriers to access reactive O-atoms. Our results provide insights for understanding CH4 oxidation on IrO2(110) surfaces under reaction conditions at which Oot atoms and adsorbed CH4 can co-exist.

Addressing the intrinsic charge transport limitation of metal oxides has been of significance for pursuing viable PEC water splitting photoelectrodes. Growing a photoelectrode with conductive nanoobjects embedded in the matrix is promising for enhanced charge transport but remains a challenge technically. We herein show a strategy of embedding laser generated nanocrystals in BiVO 4 photoanode matrix, which achieves photocurrent densities of up to 5.15 mA cm −2 at 1.23 V RHE (from original 4.01 mA cm −2 ) for a single photoanode configuration, and 6.22 mA cm −2 at 1.23 V RHE for a dual configuration. The enhanced performance by such embedding is found universal owing to the typical features of laser synthesis and processing of colloids (LSPC) for producing ligand free nanocrystals in desired solvents. This study provides an alternative to address the slow bulk charge transport that bothers most metal oxides, and thus is significant for boosting their PEC water splitting performance.