10 xLSTM Models

It’s fascinating to explore various architectures and types of models to see the whole range of AI. Transformers, as the backbone of the most powerful AI models, receive the most part of attention. However, many other AI systems, which excel in areas where Transformers meet limitations, stay overshadowed.

Today, we want to explore different options for the Extended Long Short-Term Memory (xLSTM) model. It’s a type of neural network, build on classic LSTM, used for processing and predicting sequences of data, like text or time-series information. xLSTMs have a unique structure with "memory cells" and "gates" that help them remember important information over long periods and forget irrelevant data. That’s why they remain irreplaceable in tasks like predicting text on smartphones, traffic prediction in navigation apps, music generation and recommendation.

Here are 10 modifications of xLSTM models, enhanced for implementation in various tasks:

1. xLSTM, the original version, is built on basic LSTM neural network by adding new memory and gating techniques. It matches or even exceeds Transformer performance in large-scale tasks. → Read more

   In the article “Topic 8: What are LSTM and xLSTM?” we review what we know about the origins and working process of LSTM and investigate the details of how xLSTM works. → Read more

2. Bio-xLSTM model is adapted for biology and chemistry and can process DNA, protein, and chemical data. Supporting in-context learning, it can also make protein and molecule research. → Read more

3. A Large Recurrent Action Model (LRAM), built on xLSTM, runs faster than Transformers in real-time and adapts well to longer sequences, which is crucial for areas like robotics. → Read more

4. xLSTM-Mixer combines information from time sequences and multiple data sources to provide reliable forecasts. Since the xLSTM-Mixer can handle complex data dynamics, it excels in long-term predictions. → Read more

5. Audio xLSTM (AxLSTM) is designed for learning audio features from masked spectrogram patches without labeled data. AxLSTM performs up to 20% better than similar Transformer-based models, using up to 45% fewer parameters. → Read more

6. xLSTMTime is tailored for long-term time series forecasting (LTSF). It achieves better results on various datasets, suggesting that it’s a strong alternative to Transformers in this area. → Read more

7. Vision-LSTM (ViL) adapts xLSTM for visual tasks. ViL uses stacked xLSTM blocks, alternating between processing visual patches from top to bottom and bottom to top. ViL could become a strong, versatile option for computer vision models. → Read more

8. UVixLSTM, combines CNNs (Convolutional Neural Networks) with Vision-xLSTM blocks that capture both global and temporal relationships in image patches. UVixLSTM can be used for medical image segmentation, achieving excellent segmentation accuracy with lower computational needs, and outperforming top models on the Synapse medical imaging dataset. → Read more

9. xLSTM-UNet is another option for various biomedical segmentation tasks. It’s a UNet-style model that uses Vision-LSTM that captures long-range dependencies. xLSTM-UNet provides a comprehensive solution, outperforming top CNN, Transformer, and Mamba models. → Read more

10. Seg-LSTM architecture is made to test Vision-LSTM capabilities in image segmentation tasks, suggesting areas for future improvements of the model. → Read more