Energy Efficient In-memory Hyperdimensional Encoding for Spatio-temporal Signal Processing

Geethan Karunaratne,Manuel Le Gallo,Michael Hersche,Giovanni Cherubini,Luca Benini,Abu Sebastian,Abbas Rahimi

DOI: https://doi.org/10.1109/TCSII.2021.3068126

2021-06-22

Abstract:The emerging brain-inspired computing paradigm known as hyperdimensional computing (HDC) has been proven to provide a lightweight learning framework for various cognitive tasks compared to the widely used deep learning-based approaches. Spatio-temporal (ST) signal processing, which encompasses biosignals such as electromyography (EMG) and electroencephalography (EEG), is one family of applications that could benefit from an HDC-based learning framework. At the core of HDC lie manipulations and comparisons of large bit patterns, which are inherently ill-suited to conventional computing platforms based on the von-Neumann architecture. In this work, we propose an architecture for ST signal processing within the HDC framework using predominantly in-memory compute arrays. In particular, we introduce a methodology for the in-memory hyperdimensional encoding of ST data to be used together with an in-memory associative search module. We show that the in-memory HDC encoder for ST signals offers at least 1.80x energy efficiency gains, 3.36x area gains, as well as 9.74x throughput gains compared with a dedicated digital hardware implementation. At the same time it achieves a peak classification accuracy within 0.04% of that of the baseline HDC framework.

Emerging Technologies

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to achieve high - efficiency hyperdimensional computing (HDC) in spatio - temporal (ST) signal processing. Specifically, the paper proposes an in - memory computing - based method for hyperdimensional encoding of ST signals, aiming to improve energy efficiency, reduce area occupation, and increase throughput while maintaining classification accuracy similar to that of traditional digital hardware implementations.

### Background and Objectives of the Paper

1. **Background**:
- In recent years, most breakthroughs in the field of artificial intelligence rely on complex machine - learning models, which consume a large amount of energy in training and inference tasks.
- In contrast, the human brain consumes much less energy when performing cognitive tasks, which has inspired researchers to explore brain - like computing methods.
- Hyperdimensional computing (HDC) is a brain - like computing method that represents entities by hypervectors and uses algebraic operations for combination and comparison, and is suitable for various cognitive tasks.

2. **Objectives**:
- **Improve energy efficiency**: Reduce the energy consumption of data transmission by performing calculations in memory.
- **Reduce area occupation**: Reduce the hardware area by optimizing the in - memory computing architecture.
- **Increase throughput**: Increase the data processing speed through parallel processing and optimization algorithms.
- **Maintain classification accuracy**: Ensure that classification accuracy is not sacrificed while improving performance.

### Main Contributions

1. **Proposed a new in - memory computing architecture**:
- This architecture uses a phase - change memory (PCM) cross - bar array to achieve hyperdimensional encoding of ST signals.
- By pre - computing channel - bound hypervectors and performing time encoding in memory, the need for intermediate buffers is reduced.

2. **Optimized the HDC algorithm**:
- Move part of the spatial encoding operations after time encoding to retain more spatial information, thereby improving the classification accuracy for small N - gram sizes.

3. **Experimental verification**:
- Experiments were carried out using the EMG hand - motion recognition data set. The results show that the proposed in - memory computing ST encoder has a classification accuracy close to that of the baseline model, and is significantly superior to traditional digital hardware implementations in terms of energy efficiency, area occupation, and throughput.

### Experimental Results

- **Classification accuracy**:
- The proposed in - memory computing ST encoder reaches a peak classification accuracy of 98.9% when N = 9 and L = 15, which is only 0.04% lower than the baseline model.
- Compared with the reference encoder, the classification accuracy is improved by 1.1%.

- **Energy efficiency**:
- The energy efficiency of the in - memory computing ST encoder can be as high as 75.1M N - grams/s/W, which is 1.80 to 8.83 times higher than that of a digital CMOS encoder with the same configuration.
- Compared with the encoder on a low - power general - purpose computing platform, the energy efficiency is improved by 284 times.

- **Area and throughput**:
- The total area ranges from 0.37 mm² to 0.51 mm², which is 3.36 to 6.97 times less than that of a digital CMOS encoder.
- The throughputs are 31.5M, 18.9M, and 10.5M N - grams/s respectively, which is a 9.74 - fold improvement over digital hardware implementations.

### Conclusion

This paper successfully demonstrates how to use in - memory computing technology to achieve efficient hyperdimensional encoding of ST signals on a phase - change memory cross - bar array. By optimizing the algorithm and hardware architecture, not only the energy efficiency and throughput are improved, but also a high classification accuracy is maintained, providing new ideas for future brain - like computing applications.