Class I histone deacetylases inhibition reverses memory impairment induced by acute stress in mice.

While chronic stress induces learning and memory impairments, acute stress may facilitate or prevent memory consolidation depending on whether it occurs during the learning event or before it, respectively. On the other hand, it has been shown that histone acetylation regulates long-term memory formation. This study aimed to evaluate the effect of two inhibitors of class I histone deacetylases (HDACs), 4-phenylbutyrate (PB) and IN14 (100 mg/kg/day, ip for 2 days), on memory performance in mice exposed to a single 15-min forced swimming stress session. Plasma corticosterone levels were determined 30 minutes after acute swim stress in one group of mice. In another experimental series, independent groups of mice were trained in one of three different memory tasks: Object recognition test, Elevated T maze, and Buried food location test. Subsequently, the hippocampi were removed to perform ELISA assays for histone deacetylase 2 (HDAC2) expression. Acute stress induced an increase in plasma corticosterone levels, as well as hippocampal HDAC2 content, along with an impaired performance in memory tests. Moreover, PB and IN14 treatment prevented memory loss in stressed mice. These findings suggest that HDAC2 is involved in acute stress-induced cognitive impairment. None of the drugs improved memory in non-stressed animals, indicating that HDACs inhibitors are not cognitive boosters, but rather potentially useful drugs for mitigating memory deficits.

Developing an Improved Cycle Architecture for AI-Based Generation of New Structures Aimed at Drug Discovery.

Drug discovery involves a crucial step of optimizing molecules with the desired structural groups. In the domain of computer-aided drug discovery, deep learning has emerged as a prominent technique in molecular modeling. Deep generative models, based on deep learning, play a crucial role in generating novel molecules when optimizing molecules. However, many existing molecular generative models have limitations as they solely process input information in a forward way. To overcome this limitation, we propose an improved generative model called BD-CycleGAN, which incorporates BiLSTM (bidirectional long short-term memory) and Mol-CycleGAN (molecular cycle generative adversarial network) to preserve the information of molecular input. To evaluate the proposed model, we assess its performance by analyzing the structural distribution and evaluation matrices of generated molecules in the process of structural transformation. The results demonstrate that the BD-CycleGAN model achieves a higher success rate and exhibits increased diversity in molecular generation. Furthermore, we demonstrate its application in molecular docking, where it successfully increases the docking score for the generated molecules. The proposed BD-CycleGAN architecture harnesses the power of deep learning to facilitate the generation of molecules with desired structural features, thus offering promising advancements in the field of drug discovery processes.

Octopamine integrates the status of internal energy supply into the formation of food-related memories.

The brain regulates food intake in response to internal energy demands and food availability. However, can internal energy storage influence the type of memory that is formed? We show that the duration of starvation determines whether Drosophila melanogaster forms appetitive short-term or longer-lasting intermediate memories. The internal glycogen storage in the muscles and adipose tissue influences how intensely sucrose-associated information is stored. Insulin-like signaling in octopaminergic reward neurons integrates internal energy storage into memory formation. Octopamine, in turn, suppresses the formation of long-term memory. Octopamine is not required for short-term memory because octopamine-deficient mutants can form appetitive short-term memory for sucrose and to other nutrients depending on the internal energy status. The reduced positive reinforcing effect of sucrose at high internal glycogen levels, combined with the increased stability of food-related memories due to prolonged periods of starvation, could lead to increased food intake.

Deciding what and how much to eat is a complex biological process which involves balancing many types of information such as the levels of internal energy storage, the amount of food previously available in the environment, the perceived value of certain food items, and how these are remembered. At the molecular level, food contains carbohydrates that are broken down to produce glucose, which is then delivered to cells under the control of a hormone called insulin. There, glucose molecules are either immediately used or stored as glycogen until needed. Insulin signalling is also known to interact with the brain's decision-making systems that control eating behaviors; however, how our brains balance food intake with energy storage is poorly understood. Berger et al. set out to investigate this question using fruit flies as an experimental model. These insects also produce insulin-like molecules which help to relay information about glycogen levels to the brain's decision-making system. In particular, these signals reach a population of neurons that produce a messenger known as octopamine similar to the human noradrenaline, which helps regulate how much the flies find consuming certain types of foods rewarding. Berger et al. were able to investigate the role of octopamine in helping to integrate information about internal and external resource levels, memory formation and the evaluation of different food types. When the insects were fed normally, increased glycogen levels led to foods rich in carbohydrates being rated as less rewarding by the decision-making cells, and therefore being consumed less. Memories related to food intake were also short-lived ­ in other words, long-term 'food memory' was suppressed, re-setting the whole system after every meal. In contrast, long periods of starvation in insects with high carbohydrates resources produced a stable, long-term memory of food and hunger which persisted even after the flies had fed again. This experience also changed their food rating system, with highly nutritious foods no longer being perceived as sufficiently rewarding. As a result, the flies overate. This study sheds new light on the mechanisms our bodies may use to maintain energy reserves when food is limited. The persistence of 'food memory' after long periods of starvation may also explain why losing weight is difficult, especially during restrictive diets. In the future, Berger et al. hope that this knowledge will contribute to better strategies for weight management.

Episodic long-term memory formation during slow-wave sleep.

We are unresponsive during slow-wave sleep but continue monitoring external events for survival. Our brain wakens us when danger is imminent. If events are non-threatening, our brain might store them for later consideration to improve decision-making. To test this hypothesis, we examined whether novel vocabulary consisting of simultaneously played pseudowords and translation words are encoded/stored during sleep, and which neural-electrical events facilitate encoding/storage. An algorithm for brain-state-dependent stimulation selectively targeted word pairs to slow-wave peaks or troughs. Retrieval tests were given 12 and 36 hr later. These tests required decisions regarding the semantic category of previously sleep-played pseudowords. The sleep-played vocabulary influenced awake decision-making 36 hr later, if targeted to troughs. The words' linguistic processing raised neural complexity. The words' semantic-associative encoding was supported by increased theta power during the ensuing peak. Fast-spindle power ramped up during a second peak likely aiding consolidation. Hence, new vocabulary played during slow-wave sleep was stored and influenced decision-making days later.

Leveraging transfer learning with deep learning for crime prediction.

Crime remains a crucial concern regarding ensuring a safe and secure environment for the public. Numerous efforts have been made to predict crime, emphasizing the importance of employing deep learning approaches for precise predictions. However, sufficient crime data and resources for training state-of-the-art deep learning-based crime prediction systems pose a challenge. To address this issue, this study adopts the transfer learning paradigm. Moreover, this study fine-tunes state-of-the-art statistical and deep learning methods, including Simple Moving Averages (SMA), Weighted Moving Averages (WMA), Exponential Moving Averages (EMA), Long Short Term Memory (LSTM), Bi-directional Long Short Term Memory (BiLSTMs), and Convolutional Neural Networks and Long Short Term Memory (CNN-LSTM) for crime prediction. Primarily, this study proposed a BiLSTM based transfer learning architecture due to its high accuracy in predicting weekly and monthly crime trends. The transfer learning paradigm leverages the fine-tuned BiLSTM model to transfer crime knowledge from one neighbourhood to another. The proposed method is evaluated on Chicago, New York, and Lahore crime datasets. Experimental results demonstrate the superiority of transfer learning with BiLSTM, achieving low error values and reduced execution time. These prediction results can significantly enhance the efficiency of law enforcement agencies in controlling and preventing crime.

Astrocytes control recent and remote memory strength by affecting the recruitment of the CA1→ACC projection to engrams.

The maturation of engrams from recent to remote time points involves the recruitment of CA1 neurons projecting to the anterior cingulate cortex (CA1→ACC). Modifications of G-protein-coupled receptor pathways in CA1 astrocytes affect recent and remote recall in seemingly contradictory ways. To address this inconsistency, we manipulated these pathways in astrocytes during memory acquisition and tagged c-Fos-positive engram cells and CA1→ACC cells during recent and remote recall. The behavioral results were coupled with changes in the recruitment of CA1→ACC projection cells to the engram: Gq pathway activation in astrocytes caused enhancement of recent recall alone and was accompanied by earlier recruitment of CA1→ACC projecting cells to the engram. In contrast, Gi pathway activation in astrocytes resulted in the impairment of only remote recall, and CA1→ACC projecting cells were not recruited during remote memory. Finally, we provide a simple working model, hypothesizing that Gq and Gi pathway activation affect memory differently, by modulating the same mechanism: CA1→ACC projection.

Long-term memory consolidation of new words in children with self-limited epilepsy with centro-temporal spikes.

Accelerated long-term forgetting has been studied and demonstrated in adults with epilepsy. In contrast, the question of long-term consolidation (delays > 1 day) in children with epilepsy shows conflicting results. However, childhood is a period of life in which the encoding and long-term storage of new words is essential for the development of knowledge and learning. The aim of this study was therefore to investigate long-term memory consolidation skills in children with self-limited epilepsy with centro-temporal spikes (SeLECTS), using a paradigm exploring new words encoding skills and their long-term consolidation over one-week delay. As lexical knowledge, working memory skills and executive/attentional skills has been shown to contribute to long-term memory/new word learning, we added standardized measures of oral language and executive/attentional functions to explore the involvement of these cognitive skills in new word encoding and consolidation. The results showed that children with SeLECTS needed more repetitions to encode new words, struggled to encode the phonological forms of words, and when they finally reached the level of the typically developing children, they retained what they had learned, but didn't show improved recall skills after a one-week delay, unlike the control participants. Lexical knowledge, verbal working memory skills and phonological skills contributed to encoding and/or recall abilities, and interference sensitivity appeared to be associated with the number of phonological errors during the pseudoword encoding phase. These results are consistent with the functional model linking working memory, phonology and vocabulary in a fronto-temporo-parietal network. As SeLECTS involves perisylvian dysfunction, the associations between impaired sequence storage (phonological working memory), phonological representation storage and new word learning are not surprising. This dual impairment in both encoding and long-term consolidation may result in large learning gap between children with and without epilepsy. Whether these results indicate differences in the sleep-induced benefits required for long-term consolidation or differences in the benefits of retrieval practice between the epilepsy group and healthy children remains open. As lexical development is associated with academic achievement and comprehension, the impact of such deficits in learning new words is certainly detrimental.

Ultra short term power load forecasting based on the fusion of Seq2Seq BiLSTM and multi head attention mechanism.

Ultra-short-term power load forecasting is beneficial to improve the economic efficiency of power systems and ensure the safe and stable operation of power grids. As the volatility and randomness of loads in power systems, make it difficult to achieve accurate and reliable power load forecasting, a sequence-to-sequence based learning framework is proposed to learn feature information in different dimensions synchronously. Convolutional Neural Networks(CNN) Combined with Bidirectional Long Short Term Memory(BiLSTM) Networks is constructed in the encoder to extract the correlated timing features embedded in external factors affecting power loads. The parallel BiLSTM network is constructed in the decoder to mine the power load timing information in different regions separately. The multi-headed attention mechanism is introduced to fuse the BiLSTM hidden layer state information in different components to further highlight the key information representation. The load forecastion results in different regions are output through the fully connected layer. The model proposed in this paper has the advantage of high forecastion accuracy through the example analysis of real power load data.

The effects of mnemonic variability and spacing on memory over multiple timescales.

The memory benefit that arises from distributing learning over time rather than in consecutive sessions is one of the most robust effects in cognitive psychology. While prior work has mainly focused on repeated exposures to the same information, in the real world, mnemonic content is dynamic, with some pieces of information staying stable while others vary. Thus, open questions remain about the efficacy of the spacing effect in the face of variability in the mnemonic content. Here, in two experiments, we investigated the contributions of mnemonic variability and the timescale of spacing intervals, ranging from seconds to days, to long-term memory. For item memory, both mnemonic variability and spacing intervals were beneficial for memory; however, mnemonic variability was greater at shorter spacing intervals. In contrast, for associative memory, repetition rather than mnemonic variability was beneficial for memory, and spacing benefits only emerged in the absence of mnemonic variability. These results highlight a critical role for mnemonic variability and the timescale of spacing intervals in the spacing effect, bringing this classic memory paradigm into more ecologically valid contexts.

The fading memories of youth.

The mystery of "infantile amnesia" suggests memory works differently in the developing brain.

A lifespan study of the confidence-accuracy relation in working memory and episodic long-term memory.

The relation between an individual's memory accuracy and reported confidence in their memories can indicate self-awareness of memory strengths and weaknesses. We provide a lifespan perspective on this confidence-accuracy relation, based on two previously published experiments with 320 participants, including children aged 6-13, young adults aged 18-27, and older adults aged 65-77, across tests of working memory (WM) and long-term memory (LTM). Participants studied visual items in arrays of varying set sizes and completed item recognition tests featuring 6-point confidence ratings either immediately after studying each array (WM tests) or following a long period of study events (LTM tests). Confidence-accuracy characteristic analyses showed that accuracy improved with increasing confidence for all age groups and in both WM and LTM tests. These findings reflect a universal ability across the lifespan to use awareness of the strengths and limitations of one's memories to adjust reported confidence. Despite this age invariance in the confidence-accuracy relation, however, young children were more prone to high-confidence memory errors than other groups in tests of WM, whereas older adults were more susceptible to high-confidence false alarms in tests of LTM. Thus, although participants of all ages can assess when their memories are weaker or stronger, individuals with generally weaker memories are less adept at this confidence-accuracy calibration. Findings also speak to potential different sources of high-confidence memory errors for young children and older adults, relative to young adults. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

An acute bout of resistance exercise increases BDNF in hippocampus and restores the long-term memory of insulin-resistant rats.

A sedentary lifestyle, inadequate diet, and obesity are substantial risk factors for Type 2 diabetes mellitus (T2DM) development. A major picture of T2DM is insulin resistance (IR), which causes many impairments in brain physiology, such as increased proinflammatory state and decreased brain-derived neurotrophic factor (BDNF) concentration, hence reducing cognitive function. Physical exercise is a non-pharmacological tool for managing T2DM/IR and its complications. Thus, this study investigated the effects of IR induction and the acute effects of resistance exercise (RE) on memory, neurotrophic, and inflammatory responses in the hippocampus and prefrontal cortex of insulin-resistant rats. IR was induced by a high-fat diet and fructose-rich beverage. Insulin-resistant rats performed acute resistance exercise (IR.RE; vertical ladder climb at 50-100% of the maximum load) or rest (IR.REST; 20 min). Cognitive parameters were assessed by novel object recognition (NOR) tasks, and biochemical analyses were performed to assess BDNF concentrations and inflammatory profile in the hippocampus and prefrontal cortex. Insulin-resistant rats had 20% worse long-term memory (LTM) (p < 0.01) and lower BDNF concentration in the hippocampus (-14.6%; p < 0.05) when compared to non-insulin-resistant rats (CON). An acute bout of RE restored LTM (-9.7% pre vs. post; p > 0.05) and increased BDNF concentration in the hippocampus (9.1%; p < 0.05) of insulin-resistant rats compared to REST. Thus, an acute bout of RE can attenuate the adverse effects of IR on memory and neurotrophic factors in rats, representing a therapeutic tool to alleviate the IR impact on the brain.

W-WaveNet: A multi-site water quality prediction model incorporating adaptive graph convolution and CNN-LSTM.

Water quality prediction is of great significance in pollution control, prevention, and management. Deep learning models have been applied to water quality prediction in many recent studies. However, most existing deep learning models for water quality prediction are used for single-site data, only considering the time dependency of water quality data and ignoring the spatial correlation among multi-sites. This research defines and analyzes the non-aligned spatial correlations that exist in multi-site water quality data. Then deploy spatial-temporal graph convolution to process water quality data, which takes into account both the temporal and spatial correlation of multi-site water quality data. A multi-site water pollution prediction method called W-WaveNet is proposed that integrates adaptive graph convolution and Convolutional Neural Network, Long Short-Term Memory (CNN-LSTM). It integrates temporal and spatial models by interleaved stacking. Theoretical analysis shows that the method can deal with non-aligned spatial correlations in different time spans, which is suitable for water quality data processing. The model validates water quality data generated on two real river sections that have multiple sites. The experimental results were compared with the results of Support Vector Regression, CNN-LSTM, and Spatial-Temporal Graph Convolutional Networks (STGCN). It shows that when W-WaveNet predicts water quality over two river sections, the average Mean Absolute Error is 0.264, which is 45.2% lower than the commonly used CNN-LSTM model and 23.8% lower than the STGCN. The comparison experiments also demonstrate that W-WaveNet has a more stable performance in predicting longer sequences.

Unlocking the potential of smartphone and ambient sensors for ADL detection.

The detection of Activities of Daily Living (ADL) holds significant importance in a range of applications, including elderly care and health monitoring. Our research focuses on the relevance of ADL detection in elderly care, highlighting the importance of accurate and unobtrusive monitoring. In this paper, we present a novel approach that that leverages smartphone data as the primary source for detecting ADLs. Additionally, we investigate the possibilities offered by ambient sensors installed in smart home environments to complement the smartphone data and optimize the ADL detection. Our approach uses a Long Short-Term Memory (LSTM) model. One of the key contributions of our work is defining ADL detection as a multilabeling problem, allowing us to detect different activities that occur simultaneously. This is particularly valuable since in real-world scenarios, individuals can perform multiple activities concurrently, such as cooking while watching TV. We also made use of unlabeled data to further enhance the accuracy of our model. Performance is evaluated on a real-world collected dataset, strengthening reliability of our findings. We also made the dataset openly available for further research and analysis. Results show that utilizing smartphone data alone already yields satisfactory results, above 50% true positive rate and balanced accuracy for all activities, providing a convenient and non-intrusive method for ADL detection. However, by incorporating ambient sensors, as an additional data source, one can improve the balanced accuracy of the ADL detection by 7% and 8% of balanced accuracy and true positive rate respectively, on average.

Gold price prediction by a CNN-Bi-LSTM model along with automatic parameter tuning.

Banking and stock markets consider gold to be an important component of their economic and financial status. There are various factors that influence the gold price trend and its fluctuations. Accurate and reliable prediction of the gold price is an essential part of financial and portfolio management. Moreover, it could provide insights about potential buy and sell points in order to prevent financial damages and reduce the risk of investment. In this paper, different architectures of deep neural network (DNN) have been proposed based on long short-term memory (LSTM) and convolutional-based neural networks (CNN) as a hybrid model, along with automatic parameter tuning to increase the accuracy, coefficient of determination, of the forecasting results. An illustrative dataset from the closing gold prices for 44 years, from 1978 to 2021, is provided to demonstrate the effectiveness and feasibility of this method. The grid search technique finds the optimal set of DNNs' parameters. Furthermore, to assess the efficiency of DNN models, three statistical indices of RMSE, RMAE, and coefficient of determination (R2), were calculated for the test set. Results indicate that the proposed hybrid model (CNN-Bi-LSTM) outperforms other models in total bias, capturing extreme values and obtaining promising results. In this model, CNN is used to extract features of input dataset. Furthermore, Bi-LSTM uses CNN's outputs to predict the daily closing gold price.

Prediction model of land surface settlement deformation based on improved LSTM method: CEEMDAN-ICA-AM-LSTM (CIAL) prediction model.

The uneven settlement of the surrounding ground surface caused by subway construction is not only complicated but also liable to cause casualties and property damage, so a timely understanding of the ground settlement deformation in the subway excavation and its prediction in real time is of practical significance. Due to the complex nonlinear relationship between subway settlement deformation and numerous influencing factors, as well as the existence of a time lag effect and the influence of various factors in the process, the prediction performance and accuracy of traditional prediction methods can no longer meet industry demands. Therefore, this paper proposes a surface settlement deformation prediction model by combining noise reduction and attention mechanism (AM) with the long short-term memory (LSTM). The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and independent component analysis (ICA) methods are used to denoise the input original data and then combined with AM and LSTM for prediction to obtain the CEEMDAN-ICA-AM-LSTM (CIAL) prediction model. Taking the settlement monitoring data of the construction site of Urumqi Rail Transit Line 1 as an example for analysis reveals that the model in this paper has better effectiveness and applicability in the prediction of surface settlement deformation than multiple prediction models. The RMSE, MAE, and MAPE values of the CIAL model are 0.041, 0.033 and 0.384%; R2 is the largest; the prediction effect is the best; the prediction accuracy is the highest; and its reliability is good. The new method is effective for monitoring the safety of surface settlement deformation.

Modeling opening price spread of Shanghai Composite Index based on ARIMA-GRU/LSTM hybrid model.

In the dynamic landscape of financial markets, accurate forecasting of stock indices remains a pivotal yet challenging task, essential for investors and policymakers alike. This study is motivated by the need to enhance the precision of predicting the Shanghai Composite Index's opening price spread, a critical measure reflecting market volatility and investor sentiment. Traditional time series models like ARIMA have shown limitations in capturing the complex, nonlinear patterns inherent in stock price movements, prompting the exploration of advanced methodologies. The aim of this research is to bridge the gap in forecasting accuracy by developing a hybrid model that integrates the strengths of ARIMA with deep learning techniques, specifically Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. This novel approach leverages the ARIMA model's proficiency in linear trend analysis and the deep learning models' capability in modeling nonlinear dependencies, aiming to provide a comprehensive tool for market prediction. Utilizing a comprehensive dataset covering the period from December 20, 1990, to June 2, 2023, the study develops and assesses the efficacy of ARIMA, LSTM, GRU, ARIMA-LSTM, and ARIMA-GRU models in forecasting the Shanghai Composite Index's opening price spread. The evaluation of these models is based on key statistical metrics, including Mean Squared Error (MSE) and Mean Absolute Error (MAE), to gauge their predictive accuracy. The findings indicate that the hybrid models, ARIMA-LSTM and ARIMA-GRU, perform better in forecasting the opening price spread of the Shanghai Composite Index than their standalone counterparts. This outcome suggests that combining traditional statistical methods with advanced deep learning algorithms can enhance stock market prediction. The research contributes to the field by providing evidence of the potential benefits of integrating different modeling approaches for financial forecasting, offering insights that could inform investment strategies and financial decision-making.

Spatial transcriptomics reveal neuron-astrocyte synergy in long-term memory.

Memory encodes past experiences, thereby enabling future plans. The basolateral amygdala is a centre of salience networks that underlie emotional experiences and thus has a key role in long-term fear memory formation1. Here we used spatial and single-cell transcriptomics to illuminate the cellular and molecular architecture of the role of the basolateral amygdala in long-term memory. We identified transcriptional signatures in subpopulations of neurons and astrocytes that were memory-specific and persisted for weeks. These transcriptional signatures implicate neuropeptide and BDNF signalling, MAPK and CREB activation, ubiquitination pathways, and synaptic connectivity as key components of long-term memory. Notably, upon long-term memory formation, a neuronal subpopulation defined by increased Penk and decreased Tac expression constituted the most prominent component of the memory engram of the basolateral amygdala. These transcriptional changes were observed both with single-cell RNA sequencing and with single-molecule spatial transcriptomics in intact slices, thereby providing a rich spatial map of a memory engram. The spatial data enabled us to determine that this neuronal subpopulation interacts with adjacent astrocytes, and functional experiments show that neurons require interactions with astrocytes to encode long-term memory.

Maternal aerobic exercise decreases the effects of a perinatal Western diet on the short and long-term memory of CD1 mouse progeny.

Maternal nutrition and physical activity during pregnancy and lactation can modify offspring development. Here, we investigated the effects of maternal aerobic exercise (AE) and Western diet (WD) on brain development, cognitive flexibility, and memory of progenies. Sixteen adult female mice were assigned to AE or sedentary groups (SED) and fed a balanced diet (BD) or WD. Offspring were categorized into four groups: WD + AE, WD + SED, BD + AE, and BD + SED. The AE group showed enhanced spontaneous alternation in the T-maze test, suggesting an improvement in working memory and tasks related to cognitive flexibility. The novel object recognition (NOR) test showed that the BD + AE pups improved their absolute discrimination and discrimination index at 24 h, which suggests a delay in memory consolidation without affecting evocation. WD + SED showed poorer discrimination and recognition memory. The pups of AE mothers had better efficiency in short-term memory, whereas WD offspring showed low performance in long-term memory. Interestingly, exercise improved tasks related to cognitive flexibility, regardless of the diet. These findings indicate that maternal diet and physical activity modify offspring development and suggest that maternal AE during pregnancy could be a beneficial intervention to counteract the adverse effects of WD by improving spatial memory and cognitive flexibility in offspring.

Multimodal representation learning for tourism recommendation with two-tower architecture.

Personalized recommendation plays an important role in many online service fields. In the field of tourism recommendation, tourist attractions contain rich context and content information. These implicit features include not only text, but also images and videos. In order to make better use of these features, researchers usually introduce richer feature information or more efficient feature representation methods, but the unrestricted introduction of a large amount of feature information will undoubtedly reduce the performance of the recommendation system. We propose a novel heterogeneous multimodal representation learning method for tourism recommendation. The proposed model is based on two-tower architecture, in which the item tower handles multimodal latent features: Bidirectional Long Short-Term Memory (Bi-LSTM) is used to extract the text features of items, and an External Attention Transformer (EANet) is used to extract image features of items, and connect these feature vectors with item IDs to enrich the feature representation of items. In order to increase the expressiveness of the model, we introduce a deep fully connected stack layer to fuse multimodal feature vectors and capture the hidden relationship between them. The model is tested on the three different datasets, our model is better than the baseline models in NDCG and precision.