Deep learning presents many opportunities and challenges. Training is a good example of the latter – it can take months or longer. An Oak Ridge National Laboratory-led team is studying how quantum computing, traditional HPC, and neuromorphic computing might be used to improve deep learning and their early work suggests each has strengths that could be leveraged independently or when used in concert with the others.

The work is presented in a paper (A Study of Complex Deep Learning Networks on High Performance, Neuromorphic, and Quantum Computers), posted on arxiv.org and also summarized in a short article (Computing – Quantum deep) on the ORNL web site. “[We] evaluate deep learning models using three different computing architectures to address these problems: quantum computing to train complex topologies, high performance computing (HPC) to automatically determine network topology, and neuromorphic computing for a low-power hardware implementation.

“Our results show the feasibility of using the three architectures in tandem to address the above deep learning limitations. We show a quantum computer can find high quality values of intra-layer connections weights, in a tractable time as the complexity of the network increases; a high performance computer can find optimal layer-based topologies; and a neuromorphic computer can represent the complex topology and weights derived from the other architectures in low power memristive hardware,” write the study team led by Thomas Potok, ORN’s Computational Data Analytics Group.

Here’s a snapshot of the computational resources used or planned for use by the researchers:

• “The quantum computer we are using is a D-Wave adiabatic quantum computer located at the University of Southern California Lockheed Martin Quantum Computing Center.”
• The HPC resource is ORNL’s Titan computer with roughly 300,000 cores, and 18,000 GPUs. “Utilizing 500 nodes of Titan, the evolutionary algorithm was trained for 32 generations with 500 individuals in the population allowing us to evaluate 16,000 networks.”
• The neuromorphic system “we will use to explore the MNIST problem is a memristive implementation of the neuroscience-inspired dynamic architectures (NIDA) system. NIDA is a simple SNN model composed of integrate-and-fire neurons and synapses with delays and weights that are affected by processes similar to long-term potentiation and long-term depression in biological brains. A digital hardware implementation based on NIDA, called Dynamic Adaptive Neural Network Array (DANNA), has also been created and is currently implemented on FPGA with a digital VLSI implementation in progress.”

There’s discussion of the strengths and weaknesses for each for each of  the three architectures: Quantum computers, for example, show promise but also impose constraints because of their ‘small’ size – “We use the MNIST dataset for our experiment, due to input size limitations of current quantum computers.”

Overall the work demonstrated the possibility of using “these three architectures to solve complex deep learning networks that are currently untrainable using a von Neumann architecture,” wrote the authors. Three highlights:

• The quantum computer experiment demonstrated that “a complex neural network, i.e., one with intra-layer connections, can be successfully trained on the MNIST problem. This is a key advantage for a quantum approach and opens the possibility of training very complex networks.”
• A HPC system “can be used to take the complex networks as building blocks and compare thousands of models to find the best performing networks for a given problem.”
• The “best performing neural network and weights can be implemented into a complex network of memristors producing a low-power hardware device. This is a capability that is not feasible with a von Neumann architecture. This holds the potential to solve much more complicated problems than can currently be solved with deep learning.”

Link to full paper: https://arxiv.org/abs/1703.05364

Link to ORNL article: https://www.ornl.gov/news/computing-quantum-deep